ComfyUI Node

TS Qwen 3

Run Qwen 3 VL locally over image, video and text to caption, describe or answer.

By AlexYez·Created 2 years ago·Updated 3 days ago· 11
TS Qwen 3
  • image
  • video
  • generated_text
  • processed_image
model_namehuihui-ai/Huihui-Qwen3.5-2B-abliterated
custom_model_id
hf_token
system_presetPrompts enhance
prompt
seed42
max_new_tokens512
precisionauto
attention_modeauto
offline_modefalse
unload_after_generationfalse
enabletrue
max_image_size1024
video_max_frames16
custom_system_prompt
CategoryTS/LLM

Inputs (17)

NameTypeDefaultDescription
model_nameCOMBOhuihui-ai/Huihui-Qwen3.5-2B-abliteratedPick a model from the list. Choose 'Custom (manual)' for a third-party model.
custom_model_idSTRINGHuggingFace repo id (e.g. 'Qwen/Qwen2-VL-7B-Instruct') or a full local path.
hf_tokenSTRINGHuggingFace token (Write/Read) for downloading models. Leave empty for public models. WARNING: saved in the workflow JSON in plain text — do not share a file that contains the token.
system_presetCOMBOPrompts enhanceSystem-prompt preset. Controls the model's behavior and response style.
promptSTRINGYour request (prompt) to the model.
seedINT420–18446744073709550000Seed for reproducible generation.
max_new_tokensINT51264–8192Maximum number of tokens in the response (output length).
precisionCOMBOautoWeight precision. 'auto' picks the best option. int4/int8 require bitsandbytes to be installed.
attention_modeCOMBOautoAttention implementation. 'flash_attention_2' is faster and more memory-efficient but needs a compatible GPU.
offline_modeBOOLEANfalseForbid downloads. Use only files already present in the models/LLM folder.
unload_after_generationBOOLEANfalseUnload the model from memory right after generation. Saves VRAM but slows down repeated runs.
enableBOOLEANtrueEnable processing. When disabled, images pass through to the output unchanged.
max_image_sizeINT102464–4096Maximum image side length. Larger resolutions require more VRAM.
video_max_framesINT164–256How many video frames to feed the model. More frames = better context but higher memory use.
imageoptIMAGEInput image.
videooptIMAGEInput video stream (batch of images).
custom_system_promptoptSTRINGYour custom system prompt. Used when 'system_preset' is set to 'Your instruction'.

Outputs (2)

NameTypeDescription
generated_textSTRINGText generated by the model.
processed_imageIMAGEInput image passed through (unchanged).