Qwen3-VL-8B for ComfyUI
ComfyUI-format LoRAs for the Qwen3-VL-8B text encoder — MiniMax-H3 prompt rewriting inside the graph. They stack, so one encoder covers both.
| File | Size | |
|---|---|---|
loras/minimax-h3-prompt-rewriter-8b_lora_rank256_bf16.safetensors |
1.30 GiB | Rewrites a short request into an H3 prompt. lightx2v's adapter, same weights, ComfyUI key naming. |
loras/qwen3vl-8b-c_abliterated-v3_lora_rank1_fp32.safetensors |
3.4 MB | Reproduces c_abliterated-v3 on the stock encoder instead of a second checkpoint. Removes refusal behaviour. |
Use
LoRAs into models/loras/, plus a Qwen3-VL-8B encoder from
Comfy-Org/Qwen3-VL into
models/text_encoders/.
CLIPLoader (type: stable_diffusion) -> LoRA on CLIP -> Generate Text -> your video model
- Leave CLIPLoader's
typeon the default. It has noqwen3vlentry and doesn't need one — anything exceptideogram4/boogu/krea2/flux2gives you the plain Qwen3-VL encoder. - Use greedy decoding (
sampling_mode: off). - The rewriter needs its trained system prompt and message layout, and reference images bind to placeholders by position. ComfyUI-NynxzH3 handles both, and adds the CLIP-only LoRA loader; otherwise build the prompt yourself from upstream's template.
Credits
Qwen3-VL-8B-Instruct (Apache-2.0) · lightx2v — all rewriting behaviour is theirs, this only changes key naming; their repo declares no license and is the source of truth · prithivMLmods (Apache-2.0).
The apache-2.0 tag covers the abliteration LoRA and anything derived from Qwen3-VL-8B-Instruct.