Winnougan/ltx-2.5-w4a8-convrot-int4-convrot-Winnougan-Blessing

LTX-2.5 — W4A8 ConvRot / INT4 ConvRot

Quantized LTX-2.5 checkpoints for ComfyUI, converted with the Star Ultimate Model Converter node using per-layer W4A8 and INT4 ConvRot quantization. These are built to run large-model video/audio generation on consumer GPUs with limited VRAM.

Converted and released by Winnougan.

Files

File	Format	Base Model	Notes
diffusion_models/ltx-2.5-22b-distilled-transformer-int4_convrot.safetensors	INT4 ConvRot	LTX-2.5 22B distilled transformer	Smallest / fastest, lowest precision
diffusion_models/ltx-2.5-22b-distilled-transformer-w4a8_convrot.safetensors	W4A8 ConvRot	LTX-2.5 22B distilled transformer	Better fidelity than INT4, larger than INT4
diffusion_models/ltx-2.5-22b-dev-transformer-int4_convrot.safetensors	INT4 ConvRot	LTX-2.5 22B dev transformer	Non-distilled base model
diffusion_models/ltx-2.5-22b-dev-transformer-w4a8_convrot.safetensors	W4A8 ConvRot	LTX-2.5 22B dev transformer	Non-distilled base model
text_encoders/gemma4-12b-with-proj-ltx-2.5-int4_convrot.safetensors	INT4 ConvRot	Gemma4-12B (LTX-2.5 text encoder, with projection heads)
text_encoders/gemma4-12b-with-proj-ltx-2.5-w4a8_convrot.safetensors	W4A8 ConvRot	Gemma4-12B (LTX-2.5 text encoder, with projection heads)

Adjust the table above to match whichever files are actually present in this repo — file names should mirror what the converter node produced (it appends -<target_format> to the source filename automatically).

Which format should I use?

• W4A8 ConvRot — higher fidelity, larger file size, requires a comfy-kitchen build with AsymW4A8Int8Layout (asym_w4a8_int8) support. Recommended default if your comfy-kitchen build supports it.
• INT4 ConvRot — smallest file size and fastest load, more aggressive precision loss. Best for the tightest VRAM budgets.

Both formats require comfy-kitchen installed alongside a recent ComfyUI build with quantized-layer loader support (comfy/ops.py + comfy/quant_ops.py).

Quantization details

The main transformer (AVTransformer3DModel) and the Gemma4 text encoder were quantized layer-by-layer with the following tensors kept at full precision to preserve stability and output quality:

Transformer (diffusion_models/*):

• All norm layers (k_norm, q_norm, layer norms)
• to_gate_logits — the audio↔video cross-attention mixing gates
• All adaln_single variants (7 total: adaln_single, audio_adaln_single, audio_prompt_adaln_single, av_ca_a2v_gate_adaln_single, av_ca_audio_scale_shift_adaln_single, av_ca_v2a_gate_adaln_single, av_ca_video_scale_shift_adaln_single)
• patchify_proj / audio_patchify_proj, proj_out / audio_proj_out
• scale_shift_table and all its variants (kept in F32, matching the original checkpoint precision)

Every one of the 48 transformer blocks is quantized uniformly — including block 0 and the audio/video embeddings-connector layers — matching Lightricks' own official INT8 ConvRot reference release.

Text encoder (text_encoders/*):

• All norm layers, embed_tokens, layer_scalar, model.norm
• vision_model.* (the vision tower)
• text_embedding_projection.* (the large audio/video conditioning-aggregation projections)
• audio_projector.*, multi_modal_projector.*

Only self_attn.{q,k,v,o}_proj and mlp.{gate,up,down}_proj in each of the 48 decoder layers are quantized.

Verification status

• INT8 ConvRot blacklist (used as the basis for these INT4/W4A8 conversions) was verified by diffing Lightricks' own official -comfy-int8-convrot reference release against the BF16 source, for both the transformer and the text encoder.
• W4A8 ConvRot — the comfy-kitchen asym_w4a8_int8 loader path itself is confirmed working in ComfyUI. The blacklist used here is reused from the INT8-verified profile; it has not been checked against an official W4A8 reference release from Lightricks (none exists yet).
• INT4 ConvRot — same caveat: blacklist reused from the INT8-verified profile, not checked against a reference INT4 release.

In short: the layers being protected are known-correct at INT8 precision. Whether that exact same protection list is optimal at W4A8 or INT4's tighter precision has not been independently confirmed — if you notice quality issues (especially around audio/video sync or fine detail), that's the first place to look.

Usage (ComfyUI)

1. Place the transformer file in ComfyUI/models/diffusion_models/
2. Place the text encoder file in ComfyUI/models/text_encoders/
3. Load with the standard LTX-2.5 diffusion model / text encoder loader nodes
4. Requires comfy-kitchen with W4A8 (AsymW4A8Int8Layout) and/or INT4 ConvRot (TensorCoreConvRotW4A4Layout) support built in, depending on which file you use

License

These are derivatives of Lightricks/LTX-2.5 and are distributed under the LTX-2.x Community License Agreement. See the Acceptable Use Policy and the original model card for full terms, including commercial-use restrictions for entities with $10M+ annual revenue.

Credits

• Base model: Lightricks/LTX-2.5
• Quantization: Winnougan using Star Ultimate Model Converter for ComfyUI