Comfy Worker Release Notes

The public history of Sogni Comfy Worker releases. Each entry says whether to update, what changed for your worker, and whether it needs disk space, a setting, or a driver.

#Updating

• Installer-managed workers follow the latest image. Use Upgrade worker in the Dashboard, or run worker-restart.bat on Windows or ./worker-restart.sh on Linux. The background updater waits until the worker is idle.
• A deployment pinned to a numbered tag (for example sogni/comfy-worker-cu13:1.0.206) must move the tag by hand.
• Do not interrupt active work. Prefer the managed upgrade or wait for an idle moment.
• Two image tracks share each version: CUDA 12 for most NVIDIA cards and CUDA 13 for compatible Ada and Blackwell systems. The installer picks the right one.

dashboard.sogni.ai/fast-workers shows the version each worker runs and whether an update is available. The Windows and Linux scripts have their own installer release notes.

#Version 1.0.220

Released: September 18, 2026 PT · Update: Recommended for all workers. MiniMax H3 workers need it to serve Reference-to-Video in 2K (and 1080p).

• Reference-to-Video in 2K. MiniMax H3 two-stage reference-to-video now delivers 2688×1536 (and 1920×1088) from up to 9 reference images, 3 reference videos, and 3 audio clips, and every two-stage reference-to-video clip length renders on 32 GB cards: the worker renders the reference pass at the 768p canvas on the tier's own model, then hands the result to the FastH3 refinement stage for the larger sizes. Model IDs minimax-h3-ref2va-fp8_r2v_2stage and _r2v_balanced_2stage with the 1344×768 (or portrait) canvas.
• The 1080p and 2K reference-to-video classes go only to CUDA 13 workers that carry the FastH3 checkpoint and have at least 60 GB of host RAM; 768p keeps the 1.0.218 rules.
• Disk and settings: nothing new for workers already on 1.0.218.
• 1.0.219 was published with the same changes but never recommended: its first two-stage reference-to-video job failed at graph build, and 1.0.220 carries the fix. Move any 1.0.219 worker to 1.0.220.

#Version 1.0.218

Released: September 17, 2026 PT · Update: Recommended for all workers. MiniMax H3 workers need it for the new two-stage reference-to-video jobs.

• Oversized uploads no longer crash the worker: reference videos and images are shrunk to what the job needs before rendering, which ends the out-of-memory restarts that video-driven jobs (animate, video-to-video, reference-to-video) caused on many hosts.
• MiniMax H3 two-stage output no longer shows the warm glowing patch that could appear in small areas such as clothing.
• The same FastH3 job now renders identically whether it is the first job after startup or follows another, so results are reproducible from the same seed.
• Two-stage reference-to-video jobs go only to 1.0.218 or newer.
• On 24 GB and 32 GB cards, Standard and Balanced MiniMax H3 jobs now go only to hosts with at least 39 GB of RAM; hosts with less kept crashing on them.
• Disk and settings: nothing new for workers already on 1.0.217.

#Version 1.0.217

Released: September 14, 2026 PT · Update: Recommended for all workers. MiniMax H3 and Pixal3D workers need it for the new jobs below.

• MiniMax H3 workers can take video jobs driven by an uploaded audio track, from a photo, from first and last frames, or from audio alone; these go only to 1.0.217 or newer.
• Pixal3D workers can build a 3D model from up to four photos of an object; these jobs go only to 1.0.217 or newer.
• MiniMax H3 2K output no longer shows the faint specks and corner lines 1.0.216 left along tile seams.
• Disk and settings: a worker with the Pixal3D pack downloads one more 5.2 GB model on restart. No new settings.

#Version 1.0.216

Released: September 13, 2026 PT · Update: Recommended for all workers.

• MiniMax H3 2K output has far fewer streaks along tile seams.
• A Fast worker with less than 31 GB of RAM now stays disconnected and reports itself unhealthy so its host can replace it, instead of reconnecting over and over. Relaxed workers are unchanged.
• Disk and settings: nothing new for workers already on 1.0.215.

#Version 1.0.215

Released: September 12, 2026 PT · Update: Recommended for all workers. Replaces 1.0.214, which was never made current.

• Renders that stack several personal LoRAs go only to workers on this version or newer.
• Future changes to personal LoRA rules (strengths, trigger phrases, stack size) no longer need a worker update.
• When a model download would run out of disk, the worker deletes idle personal LoRAs to make room instead of failing.
• Jobs that need a specific model file go only to workers that already hold it.
• Disk and settings: a MiniMax H3 worker that has not restarted since September 12 downloads one more 659 MB file on restart. No new settings.

#Version 1.0.214

Released: September 11, 2026 PT · Update: Superseded by 1.0.215 before it became current. Update to 1.0.215 instead.

• Prepared workers for the full personal LoRA strength range. Its changes ship in 1.0.215.

#Version 1.0.213

Released: September 11, 2026 PT · Update: Recommended for all workers. Required for MiniMax H3 workers that will serve Balanced jobs with a custom video shift.

• MiniMax H3 Balanced runs one recipe in every mode and accepts a custom video shift. Jobs with a custom shift go only to 1.0.213 or newer.
• Disk and settings: frees about 2 GB by removing a superseded MiniMax H3 LoRA at startup. A few H3 workers download a 780 MB adapter they did not already hold. No new settings.

#Version 1.0.212

Released: September 11, 2026 PT · Update: Recommended for all workers. Required for personal LoRA jobs, FlashVSR jobs with non-default options, and LTX audio-to-video jobs.

• Workers can now serve personal LoRAs that customers upload and Sogni approves. Files that do not match the model are refused cleanly and the job is retried elsewhere.
• LTX audio-to-video now animates to the uploaded soundtrack.
• FlashVSR video upscaling accepts detail, speed and seed options. Requests that use them go only to 1.0.212 or newer.
• MiniMax H3 gains a 2K output option, rolling out on FastH3 Turbo first. Peak GPU memory stays within the base job's.
• Nosana MiniMax deployments with 31 GB or more of RAM also serve FlashVSR. Pre-mount the pack to avoid a 6.4 GB download on restart.
• A missing license file no longer blocks a workflow.
• Disk and settings: new setting Personal LoRA cache (GiB) (PERSONAL_LORA_CACHE_GB), default 25 GiB, adjustable from the Dashboard. Set it to 0 to opt out of personal LoRA jobs. MiniMax H3 workers download a 690 MB upscaler file on their next model scan.

#Version 1.0.211

Released: September 10, 2026 PT · Update: Recommended for all workers. Required for FlashVSR video upscaling.

• New job type: FlashVSR upscales finished videos to 1080p or 1440p. Served by default on NVIDIA cards with at least 23 GB of VRAM and an Ampere or newer GPU. Other workers skip it and its download.
• Disk and settings: eligible workers download about 6.4 GB on their next model scan. No new settings.

#Version 1.0.210

Released: September 8, 2026 PT · Update: Recommended for all workers. Required for standalone BiRefNet background removal.

• Background removal no longer fails on simple subjects. Those failures were also taking healthy workers out of background-removal work for an hour at a time.
• Pixal3D keeps only its prompt-free reconstruction path.
• Disk and settings: nothing new.

#Version 1.0.209

Released: September 8, 2026 PT · Update: Recommended for all workers. Required for Qwen3-TTS, standalone BiRefNet background removal, or Pixal3D. Includes the 1.0.208 MiniMax H3 fix, so hosted deployments should move directly here.

• Qwen3-TTS adds three opt-in speech workflows on CUDA 13 workers with at least 24 GB of VRAM.
• Standalone BiRefNet background removal now delivers its transparent PNG correctly.
• Pixal3D handles portrait inputs correctly. It remains CUDA 13 only, local ComfyUI only, 30 GB of VRAM or more.
• Disk and settings: nothing downloads unless you select the workflow. Qwen3-TTS is about 12.7 GB, BiRefNet about 450 MB, Pixal3D about 13.4 GB.

#Version 1.0.208

Released: September 8, 2026 PT · Update: Superseded by 1.0.209, which carries the same fix. Update directly to 1.0.209.

• MiniMax H3 LightX2V Turbo returns to its earlier acceleration LoRA, which gives better motion and fewer artifacts. LightX2V Turbo jobs go only to 1.0.208 or newer.
• Pixal3D no longer needs a text prompt.
• Disk and settings: MiniMax H3 workers swap one 1.8 GB file for another of the same size. No new settings.

#Version 1.0.207

Released: September 8, 2026 PT (beta) · Update: Evaluation only. latest stayed on 1.0.206. Pin 1.0.207 or the beta tag to try it.

• New opt-in workflow: Pixal3D turns one image into a textured 3D model. Offered only to CUDA 13 workers with at least 30 GB of VRAM whose ComfyUI runs on the same machine. Select it under Workflows this worker serves to download about 13 GB.
• A worker opted into a pack with duplicate file names no longer runs out of work and restarts.
• Z-Image Turbo accepts up to 12 steps.
• SAM 3 image segmentation no longer fails on some results.
• Disk and settings: nothing new unless you opt into Pixal3D.

#Version 1.0.206

Released: September 7, 2026 PT · Update: Recommended for all workers, especially MiniMax H3 workers. Pinned deployments should move to 1.0.206.

• MiniMax H3 Turbo returns to its earlier attention settings after reports of artifacts. Longer clips return to their earlier render time.
• Interrupted model downloads recover more reliably.
• Disk and settings: nothing new.

#Version 1.0.205

Released: September 6, 2026 PT · Update: Recommended for all workers. Pinned deployments should move directly to 1.0.205; do not pin 1.0.204.

• FastH3 renders with one user LoRA are faster on 32 GB and larger cards. One sample went from 160 to 106 seconds.
• New setting Excluded LoRAs (EXCLUDED_LORA_IDS) keeps specific LoRAs from being assigned or downloaded, and removes a file only they used once the worker is idle.
• Opt-in workflow packs are selected with the existing Workflows this worker serves setting.
• The Windows installer reports the RAM actually available to Docker and can update .wslconfig for you when run interactively. CUDA 13 selection now checks every GPU's compute capability.
• Disk and settings: nothing downloads because of this update.

#Version 1.0.204

Released: September 6, 2026 PT · Update: Superseded by 1.0.205 before it was recommended. Do not pin it.

• Published during an interrupted release. Its changes ship in 1.0.205.

#Version 1.0.203

Released: September 5, 2026 PT · Update: Recommended for all workers. Important for 1.0.193 and 1.0.194, which can no longer serve Dark Beast Krea 2. Carries everything in 1.0.201 and 1.0.202.

• MiniMax H3 Balanced: only reference-to-video keeps the community adapter. The other Balanced modes return to the LightX2V LoRA and to their 1.0.199 speed.
• Workers stop offering models whose files were retired from the catalog, and report a clear error if a file is missing, instead of failing job after job. Sogni stops sending that model to the worker.
• Corrupt model files repair themselves.
• Long MiniMax H3 Turbo clips render about 17% faster with less flicker.
• Disk and settings: re-downloads a 1.8 GB MiniMax H3 LoRA if a previous release removed it. No new settings.

#Version 1.0.202

Released: September 5, 2026 PT · Update: Superseded by 1.0.203 before it was recommended. Do not pin it.

• A startup check withheld the ACE-Step 1.5 workflows. Fixed in 1.0.203, which carries the rest of its changes.

#Version 1.0.201

Released: September 5, 2026 PT · Update: Superseded by 1.0.203 before it was recommended. Do not pin it.

• Made long MiniMax H3 Turbo clips faster and steadier. Ships in 1.0.203.

#Version 1.0.200

Released: September 4, 2026 PT · Update: Recommended for MiniMax H3 workers, and needed to keep serving H3 Balanced.

• Every MiniMax H3 Balanced mode moves to the community adapter artists preferred. Balanced renders take about a third longer at the default size.
• Disk and settings: removes two 1.8 GB LoRA files and downloads a 744 MB adapter on the next start. No new settings.

#Version 1.0.199

Released: September 4, 2026 PT · Update: Recommended for all workers. Important on RTX 4090 and RTX 5080 hosts.

• Model switches are reported with their real duration. Sogni now sends cold jobs to the workers that switch fastest and asks slow switchers to switch less often.
• RTX 4090 workers get 2 GB more VRAM for models. RTX 5080 workers get their own memory profile.
• Disk and settings: nothing new.

#Version 1.0.198

Released: September 4, 2026 PT · Update: Recommended for all workers. Apple Silicon hosts should take it.

• Image renders finish about two seconds faster when the artist has the Sensitive Content Filter off. What gets withheld is unchanged.
• Mac workers: quantized models no longer crawl. NVIDIA hosts are unaffected.
• Disk and settings: nothing new.

#Version 1.0.197

Released: September 4, 2026 PT · Update: Recommended for all workers. Important on 24 GB and larger cards.

• A worker no longer drops its connection, and the job it was holding, while loading a large model.
• Workers now stay on one model family for longer, so the faster LoRA renders from 1.0.196 stick.
• Chroma image-to-image works again.
• Mac workers no longer accept extra ComfyUI start-up arguments from the environment.
• Disk and settings: nothing new.

#Version 1.0.196

Released: September 4, 2026 PT · Update: Recommended for all workers. Worth taking promptly on 24 GB cards.

• LoRA renders on quantized models are much faster on 24 GB cards. On an RTX 4090 one test went from 36 seconds to under 9.
• RTX 3090 and RTX 3080 workers stop advertising Z-Image and Qwen Image Edit, which returned corrupt output on those cards. Everything else is unchanged and the worker is not faulty.
• A CUDA 13 worker on a driver older than R580 now holds with a clear log line instead of restarting in a loop. Use the CUDA 12 image on those hosts.
• Disk and settings: nothing new.

#Version 1.0.195

Released: September 3, 2026 PT · Update: Recommended for all workers. latest moved here directly from 1.0.193.

• Dark Beast Krea 2 moves to the 3.0 checkpoint, which fixes blown-out skin and waxy textures. Renders are slower on the CUDA 12 track than on CUDA 13.
• MiniMax H3 reference-to-video Balanced is faster at every clip length.
• LTX-2 video-to-video no longer exhausts host RAM on long uploads.
• Time estimates after a model switch are accurate.
• Disk and settings: new setting EXCLUDED_WORKFLOW_IDS excludes specific models without giving up automatic model choice, and frees the disk only they used. Downloads a 13.2 GB Dark Beast file and removes the old one.

#Version 1.0.194

Released: September 3, 2026 PT · Update: Recommended for hosted deployments. latest did not move to it.

• A worker now checks at startup whether its machine is big enough and stops if not, so the provider can move it somewhere better.
• Custom LoRAs on every FastH3 mode now run on 24 GB CUDA 13 cards, including the RTX 4090 and RTX 3090.
• Disk and settings: optional MIN_VIABLE_DISK_GB, MIN_VIABLE_RAM_GB and HEALTH_PROBE_PORT for hosted deployments. Nothing changes without them.

#Version 1.0.193

Released: September 1, 2026 PT · Update: Required for MiniMax H3 Turbo and FastH3 Turbo workers. Recommended for all.

• Fixes the visual artifacts from 1.0.191 and 1.0.192.
• FastH3 Turbo is a new option about twice as fast as LightX2V Turbo. It runs on 24 GB CUDA 13 cards, including the RTX 4090 and RTX 3090. Custom LoRAs need 32 GB. Not offered on the CUDA 12 track.
• Disk and settings: nothing new.

#Version 1.0.192

Released: September 1, 2026 PT · Update: Recommended for all workers. Required for new MiniMax H3 Turbo workers.

• Large model downloads resume after a network interruption instead of starting over, so fresh workers enter service reliably.

#Version 1.0.191

Released: September 1, 2026 PT · Update: Superseded by 1.0.193. Do not use it for MiniMax H3 Turbo.

• Its faster MiniMax H3 Turbo path produced visual artifacts. 1.0.193 corrects it.

#Version 1.0.190

Released: August 31, 2026 PT · Update: Required for new or ephemeral workers on 1.0.189. Recommended for all.

• Fixes a startup regression that could leave a new worker idle instead of downloading its models and taking jobs.
• Restores installer-managed upgrades on Windows startup, and fixes live Dashboard settings updates.

#Version 1.0.189

Released: August 31, 2026 PT · Update: Recommended for all workers.

• Hardware care: the worker pauses new jobs when the GPU reaches about 189°F and resumes near 174°F, and takes a one-minute rest after 30 minutes of actual rendering. It never interrupts an active job. Both limits are adjustable from the Dashboard, and a natural idle gap already counts as the rest.
• Supported worker settings can be managed from the Dashboard.
• Disk and settings: the render-time rest and thermal settings are new. No downloads.

#Version 1.0.188

Released: August 30, 2026 PT · Update: Superseded by 1.0.189. If you are on 1.0.186 or 1.0.187, update.

• Turns off an over-cautious check that made capable GPUs decline jobs they could complete.

#Version 1.0.187

Released: August 30, 2026 PT · Update: Superseded. Update to the current release.

• First correction for workers declining valid jobs. Still too cautious; fixed in 1.0.188.

#Version 1.0.186

Released: August 29, 2026 PT · Update: Superseded. Update to the current release.

• Better memory handling and recovery, especially on Windows, but one new check declined some valid jobs. Corrected in 1.0.187 and 1.0.188.

#Version 1.0.185

Released: August 28, 2026 PT · Update: Superseded by 1.0.189.

• Refreshed the MiniMax H3 Balanced acceleration files. Replaced files were cleaned up automatically.

#Version 1.0.184

Released: August 28, 2026 PT · Update: Superseded by 1.0.189.

• Broad reliability update for MiniMax H3: previews, recovery, monitoring and compatibility.

#Version 1.0.183

Released: August 27, 2026 PT · Update: Superseded by 1.0.189.

• Fixed a MiniMax H3 failure after a model was unloaded and reloaded, present in 1.0.181 and 1.0.182.

#Older releases

Every version below is superseded. A worker updates straight to the current release; it never needs an intermediate version.

1.0.151 and 1.0.153 have no public image.