Prompt Length & Token Limits
Every model has a limit on how much prompt it reads. Some count tokens, the pieces their text encoder splits your words into; others count characters. Sogni never cuts your prompt short: if it is over the model's limit, the render is stopped before it starts, you are not charged, and the message tells you how long your prompt is and what the model accepts.
Sogni Studio shows a counter next to the prompt box as you get close to the limit, turns it red when you pass it, and holds the Generate button until you shorten the prompt.
#What counts toward the limit
- Prompt and Style together. Your style is added to the end of your prompt (as
, <style>), so both count toward the same limit. - The negative prompt on its own. Where a model reads a negative prompt, it is checked separately against the same limit.
- The longest variant of a dynamic prompt. A prompt with
{a|b}groups is measured by the longest prompt it can expand to, since each image receives one variant, never the whole template. See Dynamic Prompts.
#Limits by model
| Model | Limit |
|---|---|
| Stable Diffusion models (SD 1.5, SD 2.x, SDXL, SDXL Turbo, LCM, SD3 and the Sogni community checkpoints) | 77 tokens on the Relaxed Supernet, 6,000 characters of prompt and of negative prompt on the Fast Supernet |
| Flux (schnell, Kontext, Krea) and Chroma | 4,096 tokens |
| Wan 2.2 (every mode, Animate Move and Replace included) | 4,096 tokens |
| MiniMax H3 | 7,000 characters |
| Wan 3 | 20,000 characters |
| GPT Image 2 and 2.5 | 32,000 characters |
| HappyHorse | 5,000 characters, each Chinese character counting twice |
| ACE-Step | 4,096 characters of caption and 4,096 of lyrics |
| MiniMax Music 3 | 5,000 tokens of caption and lyrics together |
| Speech (Qwen3-TTS) | 4,096 characters of script |
| SAM 3 selection text | 32 tokens, about 20 words |
| Every other model (Krea 2, Qwen Image and Qwen Image Edit, Z-Image, Flux.2, LTX and more) | 6,000 characters of prompt, negative prompt and lyrics |
The start and end markers a text encoder adds count toward its token limit, so a 77-token Stable Diffusion prompt leaves room for about 75 words.
#Why Stable Diffusion differs by network
Stable Diffusion reads prompts with a CLIP text encoder, one 77-token window at a time. Relaxed Supernet workers (Macs running Core ML) read a single window, so anything past 77 tokens would be dropped. Fast Supernet workers read the prompt in consecutive windows and keep all of it, so on Fast a Stable Diffusion prompt can be up to 6,000 characters. Switch to Fast for a longer Stable Diffusion prompt.
#Tokens and characters
A token is not always a word. Common words are one token; unusual words split into pieces ("reddog" becomes red + dog), and punctuation takes tokens too. English runs about four characters per token, so 4,096 tokens holds roughly 16,000 characters of English, while Chinese, Japanese and Korean can use a token per character or more. That is why token limits are counted with each model's own tokenizer rather than guessed from the length.
#Tips
- Use Prompt Enhancer for a prompt that fits. Prompt Enhancer is told the active model's limit, less the space your style takes, and a rewrite that would still be too long is flagged instead of applied.
- Frontload the important stuff. Most models weight earlier words more heavily. Subject first, then style, then details.
- Keep the style short when the prompt is long. It counts toward the same limit.
- Choose a model for a very long description. Flux and Wan 2.2 read about 16,000 characters of English, Wan 3 20,000 and GPT Image 32,000. On Stable Diffusion, use the Fast Supernet (6,000 characters).