ComfyUI Node
💬 Ray's LLM: Fast Qwen3-VL Infer
Generate text with ComfyUI-native Qwen models. Qwen3-VL supports optional images and Ray's fast decode path; text-only Qwen models ignore image input and use their native ComfyUI generator.
💬 Ray's LLM: Fast Qwen3-VL Infer
- clip
- image
- text
◄promptDescribe this image in detail.►
◄max_tokens256►
◄temperature0.00►
◄top_p0.90►
◄top_k40►
◄seed42►
◄modefast►
◄sync_interval8►
◄thinkingfalse►
◄repetition_penalty1.00►
◄presence_penalty0.0►
◄repeat_guardtrue►
◄generation_presetcustom►
Category👑 Ray/💬 LLM
Inputs (15)
| Name | Type | Default | Description |
|---|---|---|---|
| clip | CLIP | ComfyUI-native generative Qwen text encoder: Qwen3, Qwen3.5, Qwen3-VL, or another compatible Qwen wrapper. | |
| prompt | STRING | Describe this image in detail. | Prompt for the Qwen model. The connected tokenizer owns the appropriate Qwen chat template. |
| max_tokens | INT | 2561–8192 | Maximum generated tokens. Reaching the cap truncates the answer. |
| temperature | FLOAT | 0.000–2 | Sampling randomness. 0 is greedy decoding; higher values vary the wording. |
| top_p | FLOAT | 0.900.01–1 | Nucleus sampling threshold. Ignored by greedy decoding. |
| top_k | INT | 400–1000 | Top-k sampling limit. 0 disables it. Ignored by greedy decoding. |
| seed | INT | 420–9223372036854776000 | Sampling seed when temperature is above 0. |
| imageopt | IMAGE | Optional image or image batch for Qwen3-VL. If the connected Qwen model is text-only, this input is ignored. | |
| modeopt | COMBO | fast | fast: use Ray's optimized path when the model is a compatible Qwen3-VL; otherwise automatically use native ComfyUI generation. native: always use the model's stock ComfyUI generator. |
| sync_intervalopt | INT | 81–32 | Qwen3-VL fast mode only: check stop tokens every N steps. May compute up to N-1 extra tokens which are discarded. |
| thinkingopt | BOOLEAN | false | Pass the model's thinking-mode request through its tokenizer. Effect depends on the connected Qwen tokenizer/checkpoint. |
| repetition_penaltyopt | FLOAT | 1.001–2 | Scale down scores of tokens already generated. 1 disables it. |
| presence_penaltyopt | FLOAT | 0.00–2 | Discourage reusing generated tokens. Set 0 to disable. |
| repeat_guardopt | BOOLEAN | true | Qwen3-VL fast mode only: stop if an identical 8–256-token phrase repeats three times. |
| generation_presetopt | COMBO | custom | custom uses the penalty controls (neutral defaults). anti_loop explicitly selects repetition 1.0 and presence 1.5; changes model behavior. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| text | STRING | Decoded answer from the connected Qwen model. |