Qwen3.5-9B for ComfyUI

ComfyUI-format LoRAs for the Qwen3.5-9B text encoder. They stack, so one encoder covers all of them.

Use

LoRAs into models/loras/, plus a Qwen3.5-9B encoder from Comfy-Org/Qwen3.5 (qwen3.5_9b_bf16 or qwen3.5_9b_int8_convrot) into models/text_encoders/.

CLIPLoader -> Load LoRA (Text Encoder) -> Generate Text

• Leave CLIPLoader's type on the default — Qwen3.5-9B is detected from the weights.
• Strength 1.0 is the merged model exactly; the rewriters were trained for 1.0.
• Greedy decoding (sampling_mode: off), thinking off. The rewriters answer directly after an empty <think></think> block, which is what ComfyUI's Qwen3.5 template emits with thinking off.

The rewriters need their trained prompt

lightx2v ships the 9B checkpoints without a prompt template. Their Omni rewriter bundle from the same training generation documents the system prompts (system_prompt.py: SYSTEM_PROMPT for T2AV / I2AV / L2AV / FL2AV, REF2AV_SYSTEM_PROMPT for Ref2AV) and the request layout (infer.py), and both 9B LoRAs answer in the right schema with it:

<|im_start|>system
{SYSTEM_PROMPT or REF2AV_SYSTEM_PROMPT}<|im_end|>
<|im_start|>user
Ordered MiniMax-H3 references:          <- Ref2AV only; keyframe tasks label each picture
<Picture 1>:                               with its role ("— exact first frame at 0.00 seconds:")
<|vision_start|><|image_pad|><|vision_end|>

Rewrite request:
task: REF2AV                               (T2AV / I2AV / L2AV / FL2AV with the other LoRA)
resolution: 16:9
effective_duration: 10.21s                 (H3's 17n+5 frame grid at 24 fps)
raw_prompt: {your request}<|im_end|>
<|im_start|>assistant
<think>

</think>

Feed that whole string to Generate Text with use_default_template off; ComfyUI binds each <|image_pad|> to the next image in the batch, in order.

How they were made

A plain weight diff, merged − Qwen/Qwen3.5-9B, factored per matrix with SVD in fp32 from the bf16 checkpoints. Every adapted matrix has exactly 32 singular values above the bf16 rounding floor and then a cliff (lightx2v merged a rank-32 adapter with ms-swift), so rank 32 recovers the adapter itself; the residual is the merge's own rounding.

• 248 matrices, language model only: every linear in all 32 layers — linear_attn.in_proj_qkv / in_proj_z / in_proj_a / in_proj_b / out_proj, self_attn.q/k/v/o_proj, mlp.gate/up/down_proj. The vision tower, embeddings, lm_head, norms and the MTP head are byte-identical to the base.
• Delta energy kept at rank 32: 98.3% (Ref2VA) and 99.8% (FL2VA) (the rest is bf16 rounding noise of the merge).
• Keys are ComfyUI's generic text-encoder form (text_encoders.transformer.model.layers.N.<module>.lora_{down,up}.weight), bf16, no alpha.

The uncensored LoRA is documented in its earlier revision: rank 4, language model only, next-token KL to the uncensored model 0.337 → 0.0007.

Credits

Qwen3.5-9B (Apache-2.0) · lightx2v (Apache-2.0) — all rewriting behaviour is theirs, this only re-expresses their merged checkpoints as LoRAs · HauhauCS (Apache-2.0) · nDimensional (Apache-2.0).