Hugging Face's logo

Comfy-Org
/
Qwen-Image-2.1

Diffusion Single File
comfyui
Model card Files
xet
 Community
15

Thanks for the Prompt Enhancing models

#3
by Landsharkbaby - opened
Discussion
Landsharkbaby
  1. What is the difference of these two PE models? Special fine-tuning for t2i,i2i? Could we use one of it, to do PE(in different llm workflow with different system prompt)?
  2. Could we just use qwen3vl_8b_int8_convrot.safetensors to do PE(in different llm workflow with different system prompt)?
mattiflou
•
edited 3 days ago

From the original repo:

Both checkpoints are fine-tuned Qwen3.5-VL 9B, not the official base model:
stock architecture (model_type: qwen3_5, Qwen3_5ForConditionalGeneration,
hybrid linear/full attention, thinking on by default), post-trained weights.
Their chat_template.jinja and tokenizer.json are byte-identical, which is
what makes one codebase honest rather than merely convenient.

Each task has its own checkpoint and its own system prompt. They are not
interchangeable and there is no merged prompt: the answer contract is part of
what each model was trained on. Point --ckpt at one and give it that model's
prompt (via --system-prompt, or ship it as system_prompt.txt inside the
checkpoint directory and it is picked up automatically).

Pointing --ckpt at the official open-source Qwen3.5-VL 9B release will load and
generate, but it was never trained against either system prompt, so it does not
reliably emit the answer JSON -- expect parse_ok: false on most rows.

So technically you can do it but your milage may vary
Landsharkbaby

Thank you @Novmik . Based on your info I do further search.
Yes. It just like Chatgpt image 2.x , the background system fine-tuning the prompt heavily, to make a simple idea to a masterpiece.
In local, to make the same effects, same quality, QI2.1 rely heavily on PE too.

Sign up or log in to comment