Comfy Worker Release Notes

This is the canonical public history for Sogni Comfy Worker releases. Each entry starts with a plain-English recap; technical details and verified Docker image information are available when needed.

#TL;DR for worker hosts

• Most updates are routine. They improve performance, reliability, hardware monitoring, and access to supported models. Staying current helps the worker remain eligible for the widest compatible workload, but an ordinary update notice does not mean something is wrong with the machine.
• Most standard installations do not need to be reinstalled. Use Upgrade worker in the Dashboard, run worker-restart.bat on Windows, or run ./worker-restart.sh on Linux. Installer-managed workers use the latest image and check for a newer image during that start process.
• Current managed installations can also update automatically. The background updater waits for the worker to be idle before replacing it. Current Windows installations use the same update-aware startup after a normal computer reboot, unless the worker was deliberately stopped.
• Hosted or customized deployments can differ. A provider recipe or a deployment pinned to a numbered image tag may require the operator to select the new version. Follow the provider's update flow or the Dashboard's instructions.
• Avoid interrupting active work. Prefer the Dashboard's managed upgrade or wait until the worker is idle before manually restarting it.

Use dashboard.sogni.ai/fast-workers to see the version a worker is running and whether an update is available.

For changes to the Windows and Linux setup and lifecycle scripts, see the separate Fast Worker installer release notes.

#Version 1.0.213

Released: September 11, 2026 PT

Update guidance: Recommended for all workers, and required for MiniMax H3 workers that will serve Balanced jobs with a custom video shift.

MiniMax H3 Balanced runs one recipe in every mode. All six Balanced graphs now apply Larry's MiniMax H3 Turbo v4 step-600 EMA adapter at strength 1.0 through its resident-weight loader, with eight Euler/simple steps and a video shift of 6 (audio shift 3). The five FL2VA Balanced graphs (text to video, image to video from a first frame, a last frame or keyframes, and first/last frame) move from the LightX2V FL2VA 8-step LoRA, which produced unstable motion on image-to-video in Sogni's side-by-side review. Reference-to-video Balanced already used Larry's adapter and moves its default video shift from 12 to 6, the setting that reviewed best. Standard, LightX2V Turbo and FastH3 Turbo are unchanged.

Balanced accepts a video shift. A Balanced request may set the video shift from 4 to 12 in steps of 0.5; 6 is the default. Higher values spend more of the eight steps on overall layout and motion, lower values more on fine detail. Requests that keep the default still run on older workers; a custom shift is sent only to 1.0.213 or newer. The Balanced graphs also accept the ER-SDE sampler, which Sogni will offer once this version is serving.

Configuration and disk: No new settings. Larry's adapter, minimax_h3_turbo_v4_step600_ema.safetensors (779,849,816 bytes, models/loras), is already part of both Balanced download groups, so most MiniMax H3 workers hold it; any worker missing it downloads it on the next model scan. The superseded LightX2V FL2VA 8-step 768p LoRA, minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (1,956,193,000 bytes), is no longer downloaded and is removed at startup, freeing about 2 GB.

#Version 1.0.212

Released: September 11, 2026 PT

Update guidance: Recommended for all workers, and required for workers that will run personal LoRAs, serve the FlashVSR detail, speed and seed options, or animate LTX audio-to-video jobs to the uploaded soundtrack.

Workers can run reviewed personal LoRAs. A LoRA a customer imports into Sogni and that passes Sogni's review can now be assigned to your worker like a catalog LoRA. The worker keeps these files in a separate cache with its own disk budget, checks each file strictly against the loaded model before use, and refuses cleanly (with the job retried elsewhere) when a file does not match. Catalog LoRAs are unchanged.

LTX audio-to-video now animates to the uploaded soundtrack. In LTX 2.3 and LTX 2.5 audio-to-video and image-plus-audio-to-video jobs, the model was replacing the uploaded audio with its own during generation and animating to that; the output sounded right because the upload was muxed back in, but the motion did not follow it (for example, a character kept talking through a silent intro). The graphs now lock the uploaded audio in place for the whole generation, matching the official LTX wiring. No new download.

FlashVSR video upscaling accepts two preferences and a seed. Detail (Stable, the default, or Sharper) and Speed (Stable, the default, or Faster) are the two settings FlashVSR's authors expose; the seed makes a run repeatable (default 0, or random). Requests that use the defaults keep working on 1.0.211 and reproduce its output exactly; non-default requests are sent only to 1.0.212 or newer.

MiniMax H3 gains a 2K output option. Every MiniMax H3 mode can deliver twice its native canvas (1344×768 becomes 2688×1536): the worker generates at the native size, enlarges the video latent 2× with the community learned latent upscaler (Comfyui_Minimax_h3_latent_Upscaler by LBH-123-AI), re-encodes the keyframes at the doubled size, refines the result on overlapping tiles for two steps, and decodes once. Audio is untouched. The option is inert until Sogni's apps and API expose it; rollout starts with the FastH3 Turbo modes. In qualification on an RTX 4090 with FastH3 Turbo, the reference 1344×768 clip of 158 frames took 517 seconds with 2K from a cold model load against 160 seconds without it from a warm one; peak GPU memory stayed within the base job's 22 GB.

Nosana MiniMax deployments with 31 GB or more now also serve FlashVSR. Their catalog request includes the FlashVSR group from this version; pre-mount the pack through the deployment's configuration to avoid a 6.4 GB download on restart.

License files never block a workflow. A pack's license or attribution files are fetched best-effort after its model files; a missing or unreachable license only logs a warning, and a pack whose model files are present is ready. Hosted deployments that mount license files at the container root are recognised.

Configuration and disk: One new operator setting, Personal LoRA cache (GiB) (PERSONAL_LORA_CACHE_GB in .env), default 25 GiB and at most 128 files, applied live from the Dashboard. Set it to 0 to opt out of personal LoRA jobs; the worker then clears that cache once active jobs finish. Shrinking it evicts the least recently used inactive files; the existing storage limits and free-disk reserve still apply. The MiniMax H3 2K upscaler weight, minimax_h3_latent_upscaler_3d_fp16.safetensors (690,592,672 bytes, models/latent_upscale_models), is downloaded by MiniMax H3 workers on their next model scan once Sogni's catalog lists it. No other new downloads.

#Version 1.0.211

Released: September 10, 2026 PT

Update guidance: Recommended for all workers and required for workers that will serve FlashVSR video upscaling.

FlashVSR v1.1 adds standalone video upscaling. A finished video can now be sent to a worker as its own job and enlarged to 1080p or 1440p. The worker keeps every frame, the exact frame rate, the full aspect ratio and the original audio track; it does not crop, trim, interpolate or re-time the clip. Eligible NVIDIA workers on either image track serve it by default: the model needs at least 23 GB of GPU memory and an Ampere, Ada, Hopper or Blackwell GPU (compute capability 8.0 or newer). Workers that do not qualify skip the workflow and its download. In qualification, the reference 1344×768 clip of 158 frames at 24 fps upscaled to 2520×1440 in 212 seconds on an RTX 4090 and 101 seconds on an RTX 5090, both from a cold model load.

Configuration and disk: No new operator setting. Eligible workers download the FlashVSR v1.1 model pack, 6,444,979,177 bytes in four files, into models/FlashVSR-v1.1 on their next model scan. Workers under 23 GB, on remote ComfyUI, or on GPUs older than Ampere do not download it. Hosted replacements without persistent storage download it during provisioning like any other default model.

#Version 1.0.210

Released: September 8, 2026 PT

Update guidance: Recommended for all workers and required for workers serving standalone BiRefNet background removal.

Background removal no longer fails on simple subjects. The worker checked every finished image against a minimum file size, on the assumption that a very small file meant a corrupt render. A background-removal result is not an ordinary picture: it is mostly flat black and white, so it compresses far below that size while being perfectly valid. Subjects with a simple outline were therefore returned, judged too small, and discarded as corrupt. The worker now checks the result by its content instead, and accepts it at any size. Because a worker stops offering a model after three consecutive failures, this also withdrew otherwise healthy workers from background-removal work for an hour at a time, which looked like a shortage of available capacity.

Pixal3D reconstructs from a single path. The prompt-free BiRefNet graph, already the default since 1.0.208, is now the only reconstruction path, and the earlier prompt-guided variant has been removed. SAM 3 is unaffected as a standalone segmentation feature. Reconstruction behavior for the default path, including the 1024 shape resolution, is unchanged.

Configuration and disk: No new operator setting, and no new or removed model downloads. Workers already serving these workflows need no additional storage.

#Version 1.0.209

Released: September 8, 2026 PT

Update guidance: Recommended for all workers and required for workers serving Qwen3-TTS, standalone BiRefNet background removal, or Pixal3D. This release includes the required MiniMax H3 LightX2V Turbo correction from 1.0.208, so hosted MiniMax deployments should move directly to 1.0.209 rather than stopping at an older version.

Qwen3-TTS adds three speech workflows. Eligible CUDA 13 workers can opt into nine built-in studio voices, clone a voice from a short reference recording, or design a voice from a written description. The three modes support ten languages and return 24 kHz mono WAV audio. The complete optional pack is about 12.7 GiB and requires at least 24 GB of GPU memory.

Standalone BiRefNet background removal now delivers its transparent PNG correctly. The worker recognizes background removal as an image utility, returns the source-resolution RGBA output, and does not apply image-generation safety handling to the utility result. In qualification, an original 832×1216 portrait produced a same-size transparent PNG in 3.65 seconds on an RTX 5090.

Pixal3D handles portrait-oriented source masks correctly. Its image and mask resizing now preserve the same height/width layout, fixing a shape mismatch that could stop reconstruction on portrait inputs. Both the BiRefNet default and the still-available explicit SAM 3 variant completed from the original portrait during qualification. Pixal3D remains a soft-launched CUDA 13 workflow for local ComfyUI workers with at least 30 GB of GPU memory.

Configuration and disk: No existing worker downloads these optional packs unless its selected workflows require them. Qwen3-TTS adds 13,584,500,392 bytes when selected. Standalone BiRefNet adds 444,473,596 bytes. Pixal3D's current pack adds 13,400,512,715 bytes. Hosted replacements without persistent storage download their complete selected model set during provisioning.

#Version 1.0.208

Released: September 8, 2026 PT

Update guidance: Superseded by 1.0.209. The LightX2V correction in this release is included unchanged in 1.0.209; update directly to 1.0.209.

MiniMax H3 LightX2V Turbo returns to its earlier v1.0 acceleration LoRA. Controlled same-seed comparisons of LightX2V's 768p four-step v1.0 and v1.1 files showed materially better motion and fewer generation artifacts with v1.0. The worker restores minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors for text-to-video, image-to-video, last-frame, keyframe and first/last-frame Turbo graphs. The newer v1.1 file is removed from persistent model storage. The model IDs, four-step schedule, ER-SDE sampler, Simple scheduler, 6/3 video/audio shifts and user-selected LoRA behavior remain unchanged.

Pixal3D can isolate a subject without requiring a text prompt. The BiRefNet reconstruction graph becomes the default and the default shape resolution moves from 1536 to 1024, matching the default $0.30 price. The existing prompt-guided SAM 3 graph remains available as an explicit variant. A standalone BiRefNet background-removal workflow is also included for eligible opt-in workers.

Configuration and disk: No new operator setting is introduced. A worker serving MiniMax H3 downloads the restored 1.8 GiB v1.0 acceleration file and removes the same-size v1.1 file. Hosted replacements without persistent storage download the complete configured model set as usual.

#Version 1.0.207

Released: September 8, 2026 PT — beta

Update guidance: This release is published for evaluation only. Both latest and the recommended-image version still point to 1.0.206, so installer-managed workers stay on 1.0.206 and need to do nothing. A host that wants to evaluate the new 3D workflow can pin sogni/comfy-worker-cu13:1.0.207, or follow the moving beta tag that now exists on both tracks.

Workers can turn a single image into a textured 3D model. Pixal3D is a new prompt-guided image-to-3D workflow: one still, plus a short prompt naming the object to reconstruct, produces a downloadable GLB with base colour, metallic, roughness, normal and ambient-occlusion maps baked in.

Pixal3D is opt-in, and only some workers are eligible for it. Nothing changes on a worker that does not ask for it. It is offered only to a worker running the CUDA 13 image with at least 30 GB of GPU memory whose ComfyUI runs on the worker's own machine. A CUDA 12 worker, a smaller card, and a worker pointed at a ComfyUI server elsewhere on the network (COMFY_HOST set to anything but localhost) never see it offered. An eligible host opts in by selecting it in Workflows this worker serves (PREFERRED_WORKFLOW_IDS), which permits about 13 GB of additional model files to download—about 9.5 GB of that on a worker already carrying the SAM 3 pack, which shares a file with it. Clearing the selection returns the worker to its normal eligible workflows and downloads nothing extra.

A worker opted into a pack whose files share a filename no longer runs out of work and restarts. Downloaded-model presence and download plans were keyed by bare filename. Pixal3D is the first pack to ship several catalog files with the same name—three different files called LICENSE—so all three collapsed into one entry: the plan fetched one of them, the pack could never be reported complete, and a worker that had selected only that workflow had nothing left to serve and exited. Presence is now keyed by destination folder and filename together, which also keeps the Dashboard's downloaded-model indicator honest for any future pack that repeats a filename.

Z-Image Turbo renders 8 steps by default and accepts up to 12. The worker's own limits had drifted below what Sogni offers, so an 11- or 12-step request was clamped at the worker. Requests that name their own step count inside the range are unaffected, and the price of the model does not change.

SAM 3 image segmentation no longer fails when its scores arrive as bfloat16. SAM 3 runs under automatic mixed precision on this stack, so its scores and masks can come back in a format NumPy has no equivalent for. Every conversion in the mask-selection path now steps down to 32-bit float first. The failure was found on the new selection report described below, but the same conversion on the existing point-and-box selection path was equally exposed.

Configuration and disk: No new operator setting is introduced. A worker that does not opt into Pixal3D downloads nothing new and keeps its existing persistent model files. A worker that does opt in needs room for the pack before selecting it.

#Version 1.0.206

Released: September 7, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers, especially workers serving MiniMax H3. Hosted deployments pinned to a numbered tag should update to 1.0.206.

MiniMax H3 returns to its earlier attention settings after reports of artifacts. The six affected Turbo recipes again keep the first three sampling steps dense and apply sparse attention only on the final step when the clip exceeds the existing size threshold. The CUDA 13 image restores the earlier attention kernels and removes token routing. Longer affected clips return to their earlier rendering cost. This also restores the earlier kernel used by FastH3; prompts, model files, LoRA strengths and step counts remain the same.

Interrupted model downloads recover more reliably. Workers retain resumable progress, switch broken large transfers to shorter requests, and use capped retry delays. Compressible catalog assets such as license files now download with the correct byte counts while retaining size and hash verification.

SAM 3 self-tests read the correctly staged source image. Input-relative paths resolve inside the Comfy input folder, allowing source verification to use the original staged image.

LoRA preparation follows available memory and planned model reuse. Supported MiniMax H3 checkpoints can prepare stacked LoRAs when the whole model is resident on an 80 GB-class or larger GPU. Jobs whose scheduler does not plan to retain the model keep the existing deferred LoRA calculation, avoiding preparation that subsequent jobs cannot reuse.

Configuration and disk: No new operator setting or model download is required by this update. Existing persistent model files remain usable. Hosted replacements without persistent storage still download their configured model set.

#Version 1.0.205

Released: September 6, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers. Installer-managed workers receive it through the normal managed upgrade or restart path. Hosted deployments pinned to an exact image tag should move directly to 1.0.205; do not pin 1.0.204.

In plain English: FastH3 renders with one user LoRA can avoid repeating the same LoRA work on every sampling step when the GPU has enough room. The optimization is deliberately narrow: it applies to the FastH3 checkpoint on 32 GB-class and larger cards only when exactly one user LoRA is attached and enough of the model can remain on the GPU. In the production-shaped RTX 5090 qualification sample, a 768x1024, 362-frame image-to-video render fell from 160 seconds to 106 seconds, while peak GPU memory rose from 27.4 GiB to 28.6 GiB. This is a single-sample result, not a promise for every prompt or size. The optimization stays off on 24 GB cards, with multiple user LoRAs, and at large shapes where the model does not have enough resident headroom. The model, prompt, LoRA strength, sampler, scheduler, and step count are unchanged.

Sogni World generation stages can verify that workers used the intended source media. For supported target-still and transition requests, the worker hashes the exact staged input bytes before generation, rejects a receipt whose hashes do not match those inputs, and returns the completed artifact hash with the matching source or first/last-frame hashes. This lets Sogni World reject an output that is not bound to the selected object and transition frames. Ordinary image and video jobs do not request a receipt and are unchanged.

Worker hosts can exclude individual installed LoRAs without leaving a routing gap. The new Excluded LoRAs setting (EXCLUDED_LORA_IDS in .env) removes those IDs from the worker's positive loaded-LoRA inventory and sends a separate negative exclusion fence to Sogni, so jobs using a blocked LoRA are not assigned to that worker. Missing blocked LoRAs are not downloaded. A cached file used only by that catalog LoRA is removed at the next idle point; files shared with workflows or built-in acceleration/detailer paths are retained.

Optional workflow families use the existing workflow selection. Eligible opt-in workflows appear in the same Workflows this worker serves setting backed by PREFERRED_WORKFLOW_IDS. Selecting one permits its model pack to download; clearing the selection returns to all normal eligible workflows without adding opt-in families. There is no second optional-workflow environment variable.

Windows and CUDA image selection are more predictable. The Windows installer reports the RAM actually available to Docker/WSL2 and can offer an explicit, interactive .wslconfig memory update while preserving the rest of that file; automatic/noninteractive startup only reports advice and never rewrites it. On Linux and Windows, CUDA 13 selection now uses every GPU's reported compute capability rather than product-name or memory-suffix guesses: x86-64, R580-or-newer driver, and compute capability 8.0 or newer on every GPU are required.

Configuration and disk: No model is downloaded automatically just because of this update. Excluding a LoRA can reclaim its catalog-exclusive cached file. The CUDA 13 image contains the pinned runtime foundation for an upcoming internal SAM 3 image-segmentation workflow, but the production workflow catalog does not offer that model pack in this release; there is no operator setting or checkpoint to provision for it yet.

#Version 1.0.204

Released: September 6, 2026 PT

Update guidance: Superseded by 1.0.205 before it was ever recommended. The numbered images were published during an interrupted release, but neither latest nor the recommended-image API moved to them. Use 1.0.205, which contains the complete reconciled release; do not pin 1.0.204.

#Version 1.0.203

Released: September 5, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers, and important for anyone still on 1.0.193 or 1.0.194: those images can no longer serve Dark Beast Krea 2 correctly (see below), and Sogni now routes those models only to 1.0.195 or newer. No operator action is needed; the installer's managed upgrade or a normal restart picks it up. This release also carries everything in 1.0.201 and 1.0.202, neither of which was ever recommended (1.0.202's startup check withheld the ACE-Step 1.5 workflows on a folder-name mismatch, caught in its canary and fixed here).

In plain English: Only reference-to-video keeps the community Balanced adapter; the other MiniMax H3 Balanced modes go back to the LightX2V LoRA. 1.0.200 moved every H3 Balanced mode to Larry's community adapter, but the artist feedback that prompted it concerned reference-to-video alone. Text-to-video, image-to-video (first frame, last frame, keyframes) and first-and-last-frame Balanced return to LightX2V's official 8-step LoRA, byte-for-byte the recipe 1.0.199 ran; reference-to-video Balanced stays on Larry's adapter. Nothing changes for artists except the look, and Balanced renders on those modes return to their 1.0.199 speed.

Workers stop offering models they cannot actually run, and say why when a file is missing. When Sogni retires a model file from the catalog while an older worker's graph still names it, that worker used to keep advertising the model and fail every job for it with an unreadable validation dump. Two changes: the worker now checks its graphs against the catalog at startup and withholds only the affected models (with a log line naming the workflow, node and file) while serving everything else; and if a job still reaches a loader whose file is not installed, the job error now reads Missing model file on this worker: <folder>/<file> (<node>) with what to do about it, and Sogni stops routing that model to that worker for the rest of its connection instead of retrying it. Workers on 1.0.193 and 1.0.194 hit exactly this with Dark Beast Krea 2 after the v3.0 checkpoint replaced the original file; updating clears it.

Corrupt model files repair themselves. A checkpoint whose hash no longer matches its catalog record is re-fetched as a verified replacement with the damaged bytes kept in a recoverable quarantine, including on cached-only hosted workers. Workers also report their effective host and container memory limits precisely, so job routing sees the RAM the worker really has.

MiniMax H3 Turbo clips render faster and steadier (the 1.0.201 change, shipping here). The sparse-attention speed-up that arrived in 1.0.190 was only reaching the last of the four denoising steps on the LightX2V Turbo and reference-to-video Turbo recipes; the schedule is corrected so it applies to every step after the first, and the CUDA 13 image moves to the newer attention kernels from upstream. On an RTX 5090 a 15-second 1344x768 text-to-video clip that took about 280 seconds now takes about 232 seconds, with visibly less frame-to-frame brightness flicker. Only Turbo clips of about 197 frames or more at 1344x768 are affected, because the sparse path engages above a 60,000-token floor; shorter Turbo clips are unchanged, as are the Balanced and Standard tiers. One side effect of the newer CUDA 13 kernels: the FastVideo Turbo recipes (the explicit FastH3 Turbo ids) render slightly differently, so an artist re-running an old seed there will get a valid but not bit-identical clip. 1.0.201 carried this alone and was never recommended; its production canary passed, and a same-card comparison showed 1.0.201 renders the LightX2V Turbo recipes bit-for-bit identical to 1.0.200 at 124 frames.

Disk: workers download the 1.8 GB LightX2V FL2VA 8-step LoRA again if a previous release removed it; Larry's 744 MB adapter stays.

#Version 1.0.202

Released: September 5, 2026 PT

Update guidance: Superseded by 1.0.203 before it was ever recommended. Its numbered images were published, but a startup check in this build withheld the ACE-Step 1.5 workflows on a folder-name mismatch, caught in its canary; neither latest nor the recommended-image API moved to it. Its changes ship in 1.0.203; do not pin this tag.

#Version 1.0.201

Released: September 5, 2026 PT

Update guidance: Superseded by 1.0.203 before it was ever recommended. 1.0.201 was published and passed its production canary, but latest and the recommended-image version were never moved to it while a separate MiniMax H3 Turbo output report was investigated. Its changes ship in 1.0.203; do not pin this tag.

In plain English: MiniMax H3 Turbo clips render faster and steadier. The sparse-attention speed-up that shipped in 1.0.190 was only reaching the last of the four denoising steps on the LightX2V Turbo and reference-to-video Turbo recipes; this release corrects the schedule so it applies to every step after the first, and moves the CUDA 13 image to the newer attention kernels from upstream. On an RTX 5090 a 15-second 1344x768 text-to-video clip that took about 280 seconds now takes about 232 seconds, with visibly less frame-to-frame brightness flicker. The 4-step FastVideo recipes and the Balanced and Standard tiers are unchanged.

#Version 1.0.200

Released: September 4, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers that serve MiniMax H3, and needed to keep serving the H3 Balanced tiers as the catalog moves. No operator action is needed: the installer's managed upgrade or a normal restart picks it up, and the worker downloads the one new file (about 744 MB) on its next start.

In plain English: The MiniMax H3 Balanced tiers go back to the community adapter people preferred. Every H3 Balanced mode — text-to-video, image-to-video (first frame, last frame, and keyframes), first-and-last-frame, and reference-to-video — now runs on Larry's MiniMax H3 Turbo v4 step-600 EMA adapter instead of LightX2V's official 8-step LoRAs. Reference-to-video ran this adapter from 1.0.185 through 1.0.194; after the 1.0.195 switch to the LightX2V release, artists told us they preferred the earlier look, so this release restores it and applies the same recipe to the other Balanced modes for consistency. Step count, sampler, sizes, and model IDs are unchanged, so nothing changes for artists except the look. Balanced renders take somewhat longer per job than they did on the LightX2V files; the difference is on the order of a third at the default size.

Disk: the worker removes both LightX2V 8-step 768p LoRA files (about 1.8 GB each) from its models folder and downloads the 744 MB adapter in their place.

#Version 1.0.199

Released: September 4, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers, and important on RTX 4090 and RTX 5080 hosts, which get more of their VRAM back for models. No operator action is needed; the installer's managed upgrade or a normal restart picks it up.

In plain English: A model switch is now reported as what it is. When a job needs a model your worker does not have loaded, the previous model has to leave the GPU and the new one has to come in — ten to twenty seconds on most cards, and past a minute on a card that was full from a video model. Until now the worker told Sogni the job had started the moment the request was handed to ComfyUI, before any of that loading had happened. Artists watched a render sit at 0%, and the network measured the switch as if it were free. The worker now reports the unload and the load as separate phases with their real durations, and reports the job as started only when the first step actually begins. The web app shows "Unloading previous model" and "Loading model" while it happens, and Sogni's job routing now knows each card's real switch cost, so cold jobs go to the workers that switch fastest and hosts that switch slowly are asked to switch less often.

More VRAM for models on RTX 4090 and RTX 5080 hosts. The RTX 4090 profile returns to a 2 GB driver reserve (it had been 4 GB), giving models 2 GB more room to stay resident between jobs. The RTX 5080 gets its own profile with a 1.5 GB reserve instead of inheriting the RTX 5090's 2 GB. Other cards are unchanged.

#Version 1.0.198

Released: September 4, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers. Image renders finish faster on every job whose artist had the Sensitive Content Filter switched off. Apple Silicon hosts should take it: it is the difference between a quantized model running and effectively not running at all.

In plain English: Image renders get roughly two seconds faster on a large share of jobs. Every finished image was being passed through a content classifier — a small neural network that loads its own model, shrinks the image and runs a forward pass on the CPU — so the result could carry a content label. For an artist who asked for the Sensitive Content Filter, that check is what decides whether the render is delivered, so it earns its keep. For an artist who had switched the filter off, it decided nothing: the image was delivered either way and the label was only advisory. Those renders were paying two seconds or more each for a signal nobody acted on.

The classifier now runs only when the artist actually asked for the filter. Nothing about withholding changes — a job with the filter on is still checked and still withheld on a hit, and video and upscale jobs stay exempt from the classifier exactly as before. Results still carry a content label in every case; for a filter-off job that label now comes from the prompt-text scan alone, which is a pattern match over text the worker already holds and costs nothing measurable.

One change for Apple Silicon hosts: quantized models no longer crawl on a Mac worker. Models stored in an 8-bit format were having their maths done on the CPU, because macOS has no Metal implementation of that operation and the framework quietly fell back — once per layer, on every step. A MiniMax H3 render made no measurable progress in ten minutes. The worker now unpacks those weights to a format Metal can handle, so the work stays on the GPU while the weights stay compact in memory. NVIDIA hosts are unaffected and pick exactly the same operations they picked before.

#Version 1.0.197

Released: September 4, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers, and important on 24 GB and larger cards, where it completes the LoRA speed-up shipped in 1.0.196.

In plain English: This release finishes the work 1.0.196 started. That release made LoRA renders much faster by keeping the model resident on the card, but it also meant that changing models became a heavier operation — the whole model has to be loaded rather than streamed. Sogni has since changed how work is handed out so that a worker stays on the same family of models for longer, which makes those loads rare. Your worker will now hold one kind of model for a while instead of hopping between them, and LoRA image renders stay fast because of it.

The fix in this release is what makes that safe. A worker no longer loses its connection while loading a large model. The connection check allowed 45 seconds of quiet, which a full model load on a busy card could exceed — the worker was disconnected mid-load, the job it was holding was lost, and the worker restarted itself. That window is now two minutes, comfortably longer than any normal load, so a slow load is just a slow load.

Also in this release: image-to-image on Chroma works again with its own default sampler setting, and video models are now kept off the resident path by name rather than by a size estimate, which is more reliable.

One change for Apple Silicon hosts: the Mac worker no longer accepts extra ComfyUI start-up arguments from the environment. These decide the attention backend and the memory safety limits, and Sogni's time estimates assume the ones the worker chooses, so an override could quietly slow down or destabilise a machine. The Docker images already ignored them; both platforms now behave the same way. No action is needed unless you were setting them deliberately.

#Version 1.0.196

Released: September 4, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers, and worth taking promptly on 24 GB cards, where image renders using a LoRA get materially faster.

In plain English: Two changes matter to people running workers. Renders that use a LoRA on a quantized model are much faster on 24 GB cards. Until now the worker played it safe and streamed the model through the card layer by layer whenever a LoRA was attached, even on cards with room to spare. It now keeps the model resident where it fits, which on an RTX 4090 took a Dark Beast Krea 2 render with four LoRAs from 36 seconds to under 9 seconds in a controlled test. Video models deliberately keep the old behaviour, because they need that headroom for the render itself and gained almost nothing from the change.

Second, a worker will no longer offer a model its GPU cannot render correctly. Two models returned corrupted images on one specific GPU generation, and because the images came back looking like ordinary failures the worker kept being handed the same work. Affected machines now simply stop advertising those two models and keep serving everything else. If you run one of these cards you will see two fewer models in your list; nothing else changes, and your worker is not faulty.

#Version 1.0.195

Released: September 3, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers. Installer-managed workers on the latest tag update automatically; 1.0.194 was published but never promoted to latest, so latest moves from 1.0.193 directly to this release.

In plain English: Three changes matter to people running workers. Dark Beast Krea 2 now uses the Dark Beast 3.0 int8 checkpoint, which fixes the blown-out skin and waxy textures seen from the previous file; the old file is removed from disk automatically and the new 13.2 GB file downloads through the normal model plan. MiniMax H3 Ref2VA Balanced now uses LightX2V's official 8-step LoRA, applied so the model stays resident on the card, which is faster than before at every clip length we measured. And operators can now exclude specific models with EXCLUDED_WORKFLOW_IDS without giving up automatic model choice for everything else; excluding a model also frees the disk space only that model needed.

#Version 1.0.194

Released: September 3, 2026 PT

Update guidance: Recommended for hosted and decentralised deployments, especially on providers that place a worker on an arbitrary node. Installer-managed workers on the latest tag do not need to act; they remain on 1.0.193 until the next release.

In plain English: Two changes matter to people running workers. First, a worker now checks at startup whether the machine it landed on is actually big enough to be useful, and stops itself if it is not, so the provider can move it somewhere better instead of leaving it parked and idle. Second, custom LoRAs on every explicit FastH3 mode are now qualified on 24 GB CUDA 13 cards, including the NVIDIA GeForce RTX 4090 and RTX 3090, where they previously needed 32 GB. This release also adds the vh5tape worn-VHS LoRA for MiniMax H3 to the pre-download set, and lets renders carry a sensitivity label so apps can blur or reveal them — what gets withheld is unchanged.

#Version 1.0.193

Released: September 1, 2026 PT

Update guidance: Required for workers serving MiniMax H3 Turbo or FastH3 Turbo; recommended for all Sogni Comfy Workers.

In plain English: This release fixes the visual artifacts seen in the initial faster MiniMax H3 rollout and introduces FastH3 Turbo as a distinct, production-qualified option. It runs FastVideo's FastH3 4-step Preview v1 VSA DataFree checkpoint, renders about twice as fast as the familiar LightX2V Turbo (up to 6× faster than standard H3 on a 15-second clip), and brings accelerated H3 generation to 24 GB-class cards, including the NVIDIA GeForce RTX 4090 and RTX 3090, when at least 23 GB of usable VRAM is available. Custom LoRAs remain supported on workers with at least 32 GB of VRAM. The existing Turbo option keeps its established LightX2V look, and the selected model always determines the engine.

#Version 1.0.192

Released: September 1, 2026 PT

Update guidance: Required for workers being provisioned or updated for MiniMax H3 Turbo; recommended for all Sogni Comfy Workers.

In plain English: Large model downloads now continue from the bytes already saved after an ordinary network or CDN interruption instead of starting over. This lets fresh and ephemeral workers finish downloading the MiniMax H3 assets and enter service reliably, even when a multi-gigabyte transfer is interrupted.

#Version 1.0.191

Released: September 1, 2026 PT

Update guidance: Superseded by 1.0.193. Do not use this version for MiniMax H3 Turbo.

In plain English: This release attempted to make MiniMax H3 Turbo about twice as fast, but its FastVideo graph produced visual artifacts in production. Version 1.0.193 corrects the graph and exposes FastH3 Turbo and the familiar LightX2V Turbo as separate choices. Update to 1.0.193 instead.

#Version 1.0.190

Released: August 31, 2026 PT

Update guidance: Required for newly provisioned or ephemeral Comfy Workers running 1.0.189; recommended for all Sogni Comfy Workers.

In plain English: This release fixes a startup regression that could leave a new worker idle instead of downloading its preferred models and accepting jobs. The worker now waits until at least one regular image or video workflow is actually available before it connects, while still reporting newly downloaded models to Sogni as they become ready. Existing workers with persistent model storage were less likely to encounter the problem, but should update normally.

#Version 1.0.189

Released: August 31, 2026 PT

Update guidance: Recommended for all Sogni Comfy Workers.

In plain English: This release adds practical hardware-care controls and lets supported settings be managed from the Dashboard. It never stops work already in progress. By default, the one-minute rest is needed only after 30 minutes of actual rendering, and a natural gap of at least one minute already counts. You can adjust both the render interval and rest duration from the Dashboard. With the default settings, even nonstop back-to-back jobs leave the worker available approximately 96.8% of busy time.

#Version 1.0.188

Released: August 30, 2026 PT

Update guidance: If running 1.0.186 or 1.0.187, update to the current release; otherwise this version has been superseded by 1.0.189.

In plain English: This fixed an overly cautious check that could make a capable GPU decline jobs it could complete normally. Updating to the current release is the only action needed.

#Version 1.0.187

Released: August 30, 2026 PT

Update guidance: Superseded; update to the current release at the next convenient opportunity.

In plain English: This was the first correction for workers that were declining some valid jobs, but it was still too cautious. Update to the current release; no other troubleshooting or configuration change is needed.

#Version 1.0.186

Released: August 29, 2026 PT

Update guidance: Superseded; update to the current release at the next convenient opportunity.

In plain English: This improved memory handling and recovery, especially on Windows, but one new check could be too cautious about accepting work. Later releases keep the useful protections while correcting that behavior. Updating is all that is required.

#Version 1.0.185

Released: August 28, 2026 PT

Update guidance: Superseded by 1.0.189.

In plain English: This refreshed the acceleration files used by MiniMax H3 Balanced for better qualified performance. Existing model names and app integrations continued to work, and the worker cleaned up replaced files automatically.

#Version 1.0.184

Released: August 28, 2026 PT

Update guidance: Superseded by 1.0.189.

In plain English: This was a broad reliability update for MiniMax H3 and general worker operation. It improved previews, recovery, monitoring, and compatibility across the supported H3 job types.

#Version 1.0.183

Released: August 27, 2026 PT

Update guidance: Superseded by 1.0.189.

In plain English: This fixed a MiniMax H3 video-generation error that could appear after models were unloaded and loaded again. Updating restored normal operation without requiring users to change their projects.

#Earlier August 2026 worker releases

The following releases fill the previous gap in the public history. They are all superseded. A worker can update directly to the current release; it does not need to install each intermediate version.

#H3 Balanced and recovery: versions 1.0.179–1.0.182

Version 1.0.182 — August 27, 2026 PT

In plain English: This attempted to repair a stale GPU transfer buffer after memory was reclaimed. A related MiniMax H3 reload problem remained and was fixed in 1.0.183, so do not stop on this version.

Version 1.0.181 — August 27, 2026 PT

In plain English: This restored MiniMax H3 compatibility immediately after the Balanced rollout. Versions 1.0.182 and 1.0.183 completed the follow-up recovery fixes.

Version 1.0.180 — August 27, 2026 PT

In plain English: This introduced MiniMax H3 Balanced for text-to-video, image-to-video, first/last-frame, and reference-video work. It also moved H3 Turbo to its qualified LightX2V v1.1 acceleration files while preserving user-selected LoRAs.

Version 1.0.179 — August 26, 2026 PT

In plain English: MiniMax H3 reference-video work became substantially faster, video-to-video inputs kept their intended frame rate, and worker-produced videos used a consistent high-quality encode.

#Model options and host reliability: versions 1.0.173–1.0.178

Version 1.0.178 — August 24, 2026 PT

In plain English: This maintenance release improved managed upgrades and MiniMax H3 compatibility. The installer gave active work more time to finish cleanly and correctly detected additional installed-worker layouts.

Version 1.0.177 — August 23, 2026 PT

In plain English: This is the first release eligible for MiniMax H3 reference-video jobs that use an uploaded reference video. It also made upscaling and recovery safer by avoiding an inappropriate output-classifier pass and repairing only model files proven to be incomplete.

Version 1.0.176 — August 20, 2026 PT

In plain English: Worker hosts gained access to user-attachable LoRAs across every MiniMax H3 mode. The worker also became better at recovering from startup failures and avoiding jobs becoming stuck before rendering began.

Version 1.0.175 — August 19, 2026 PT

In plain English: MiniMax H3 Turbo moved to the qualified native 768p LightX2V acceleration file for better speed and output consistency. Cleanup of retired model files was restored at the same time.

Version 1.0.174 — August 18, 2026 PT

In plain English: This important corrective release restored image- and video-based jobs affected by 1.0.173. It also added the worker health data now shown in the Dashboard—boot time, restart counts, temperature, VRAM, and utilization—and added the Krea 2 BloomGirls and Purple Grainy LoRAs.

Version 1.0.173 — August 18, 2026 PT

In plain English: This improved WSL2 memory behavior, cancellation recovery, and RTX VSR output validation. It was quickly replaced because a security-hardening change prevented valid worker-owned image and video assets from being read; 1.0.174 restored those jobs safely.

#LTX 2.5, RTX VSR, Music 3, and H3 R2V Turbo: versions 1.0.169–1.0.172

Version 1.0.172 — August 15, 2026 PT

In plain English: RTX VSR upscaling was extended through 16K where hardware and source dimensions allow, and Nosana workers could repair an unexpectedly empty model cache.

Version 1.0.171 — August 14, 2026 PT

In plain English: This is the first fully eligible release for NVIDIA RTX VSR upscaling and MiniMax H3 Reference-to-Video Turbo. It also completed the required NVIDIA NGX packaging and runtime checks so unsupported workers would not advertise RTX VSR.

Version 1.0.170 — August 14, 2026 PT

In plain English: This introduced MiniMax Music 3, the initial MiniMax H3 Reference-to-Video Turbo workflow, and the initial RTX VSR integration. Routing for H3 Reference-to-Video Turbo and RTX VSR requires 1.0.171 or later because that follow-up completed their release safeguards.

Version 1.0.169 — August 13, 2026 PT

In plain English: This introduced the LTX 2.5 family, including distilled and development workflows for text, image, audio, combined image-and-audio, and source-video inputs.

#H3 Turbo and compatibility: versions 1.0.162–1.0.168

Version 1.0.168 — August 12, 2026 PT

In plain English: MiniMax H3 reference limits and out-of-memory recovery were tightened so an incompatible request could be redirected cleanly instead of leaving the GPU in a partial state.

Version 1.0.167 — August 12, 2026 PT

In plain English: This refreshed the MiniMax H3 Turbo acceleration file, reduced GPU-memory spill when using quantized LoRAs, and added automatic CUDA 13 image selection for compatible systems.

Version 1.0.166 — August 10, 2026 PT

In plain English: This was a rapid follow-up build during the H3 end-frame and sampler rollout. It introduced no separate operator action or model-eligibility requirement and was superseded by the next routine release.

Version 1.0.165 — August 10, 2026 PT

In plain English: MiniMax H3 image-to-video gained optional end-frame guidance, and the H3 and Qwen model families gained their validated sampler choices. Older workers could continue ordinary jobs but were not selected for these new options.

Version 1.0.164 — August 8, 2026 PT

In plain English: MiniMax H3 Turbo restored ER-SDE as its dependable default sampler after the first Turbo rollout. The newer sampler choices were already available from 1.0.163.

Version 1.0.163 — August 7, 2026 PT

In plain English: This is the first release eligible for the complete MiniMax H3 Turbo text-to-video, image-to-video, and first/last-frame workflow set with validated sampler selection.

Version 1.0.162 — August 7, 2026 PT

In plain English: This introduced the first H3 Turbo fleet build and strengthened managed worker updates. It was replaced within hours by 1.0.163, which completed the validated sampler support.

#Native H3 and Identity Edit rollout: versions 1.0.154–1.0.161

Version 1.0.161 — August 7, 2026 PT

In plain English: This was a short-lived rebuild during the Sogni Identity Edit v0.3 rollout. It introduced no separate operator action and was superseded the same day.

Version 1.0.160 — August 6, 2026 PT

In plain English: This introduced Sogni Krea 2 Identity Edit v0.3 and automatically fitted MiniMax H3 keyframes to the selected video dimensions.

Version 1.0.159 — August 5, 2026 PT

In plain English: MiniMax H3 conditioning and video export became more reliable, and reconnect activity could no longer interfere while the worker was cleaning up GPU memory.

Version 1.0.158 — August 4, 2026 PT

In plain English: MiniMax H3 output was fixed at its native 24 frames per second, reference limits were enforced before rendering, and ACE-Step 1.5 XL became available on compatible 20 GB workers.

Version 1.0.157 — August 4, 2026 PT

In plain English: This is the first release eligible for MiniMax H3 Reference-to-Video-with-Audio. It added the dedicated model assets, workflow, and request limits for video-plus-audio references.

Version 1.0.156 — August 4, 2026 PT

In plain English: This refined the initial MiniMax H3 execution path and made time estimates more representative. It did not require users to recreate existing projects.

Version 1.0.155 — August 3, 2026 PT

In plain English: This is the first published worker build with correct native MiniMax H3 output. It fixed the Qwen-VL text-encoding regression that caused 1.0.154 to be withdrawn.

Version 1.0.154 — August 3, 2026 PT

In plain English: This was the initial published native MiniMax H3 and Sogni Identity Edit v0.2 build. It carried a Qwen-VL regression and was replaced within minutes by 1.0.155. Do not deliberately install it.

#Early rollout note: versions 1.0.150–1.0.153

• 1.0.150 — August 1, 2026 PT stabilized Wan Animate preprocessing and large-LoRA memory handling on RTX 4090 workers. Its CUDA 12 digest is sha256:867b8aaae5cdc8fb23423bba555b3f236be98d946b35fe07bce5a8ac63c986ce; no CUDA 13 image was published for this release.
• 1.0.151 introduced the first Sogni Krea 2 Identity Edit alpha in source, but its public image is no longer available.
• 1.0.152 was a CUDA 12-only corrective image for RTX 4090 memory allocation. Its immutable digest is sha256:7292e42a8d98a8b47464e4179f29ca8f615f36f2fa9d52324910194884d2d7b1; it is obsolete and should not be installed deliberately.
• 1.0.153 contained the first native MiniMax H3 implementation in source but was never published as a public worker image. Version 1.0.155 is the first published H3 build considered correct for routing.

#Installer and managed-update improvements

The worker image and installer are developed together, but their version numbers have not always moved in lockstep. These are the host-visible installer improvements from the same August release window:

• August 2–3 PT: made Windows and Linux worker starts, stops, and updates mutually exclusive so two lifecycle operations could not collide; removed legacy update-checker processes; and made managed starts pull the latest image after an administrator restart or Windows reboot.
• August 7–8 PT: further serialized managed lifecycle actions and made Windows update detection resilient to older image labels.
• August 10–12 PT: added automatic CUDA 13 image selection for compatible hosts, kept ARM systems away from incompatible x86 images, and preserved compatibility with older automatic-update installs.
• August 14–15 PT: restored CUDA 13 selection after the LTX 2.5 integration and migrated Windows/WSL2 installations to the NVIDIA NGX capability required for RTX VSR.
• August 24 PT: allowed a stopping worker up to 30 seconds to finish gracefully and fixed upgrades for LLM workers and older manifests without an image name.
• August 28 PT: contained managed worker services more reliably and added per-worker memory limits for multi-GPU hosts.

Most hosts receive these improvements through the current installer-managed start and update flow. A normal worker update does not require reinstalling the entire application.

Concrete historical tags remain available in the CUDA 12 tag history and CUDA 13 tag history.