NidAll/Turbo-INT8-ConvRot-for-Qwen-Image-2.1

Turbo INT8 ConvRot for Qwen-Image-2.1

A community INT8 ConvRot quantization of the transformer from Qwen/Qwen-Image-2.1-Turbo. This is a quantized derivative of the official Turbo checkpoint, not a fine-tune, acceleration LoRA, or official Qwen release.

Only the diffusion transformer is included. Use a compatible Qwen-Image-2.1 text encoder and VAE separately.

Quantization

• Source revision: 0da70ce2c362350e8a43fdd163a371783fe6228a.
• Source component: transformer (both official BF16 shards).
• Conversion tool: ComfyUI Native Quantizer.
• Storage format: native ComfyUI int8_tensorwise with convrot: true and convrot_groupsize: 256.
• 224 attention/MLP weight matrices quantized across 32 transformer blocks.
• Input/output projections, text conditioning, timestep/shared modulation and normalization retain source precision.
• Rotation: FP32 butterfly; per-layer weight reconstruction threshold: 0.02 relative L2.

Modification notice: the original transformer weights were converted to INT8 ConvRot with protected tensors retained at source precision. Publication metadata removes conversion-machine details; tensor payloads are unchanged from the tested quantized checkpoint.

ComfyUI usage

1. Put qwen-image-2.1-turbo-int8-convrot.safetensors in ComfyUI/models/diffusion_models/.
2. Select it in Load Diffusion Model, with weight_dtype: default.
3. Use the native Qwen-Image-2.1 conditioning path with its compatible text encoder and VAE.
4. Use Euler, CFG 1, and the official eight-step schedule below. No acceleration LoRA is required.

Use ManualSigmas connected to SamplerCustom (CFG 1) or SamplerCustomAdvanced with BasicGuider. Select euler through KSamplerSelect:

1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0

The terminal zero completes eight denoising steps. Do not apply an additional time shift. These values come from the official saved schedule; the Turbo scheduler configuration disables dynamic shifting. Ordinary named schedulers do not reproduce this schedule. Other schedules are not evaluated by the original authors.

Use a ComfyUI version supporting Qwen-Image-2.1 and native INT8 ConvRot. This single-file artifact is intended for ComfyUI, not direct Diffusers from_pretrained loading. Editing uses the native Qwen-Image-2.1 reference-image conditioning path; editing was not tested for this artifact.

Validation and limits

Standalone structural, quantization metadata and payload-integrity verification passed. A user-supplied 1024 x 1024 text-to-image result confirms successful generation using this checkpoint, Euler, the official schedule, CFG 1, a W4A8 text encoder and Comfy Kitchen attention. The demanding illustrated-recipe prompt produced readable labels and a coherent composition.

This is one functional example, not a controlled quality comparison with BF16. Quantization loss, broad prompt coverage, editing quality, LoRA compatibility, latency and peak memory have not been established. The checkpoint's conversion-time runtime_verified flag remains false; the separate generation example is the runtime evidence described here.

License and attribution

The original Qwen Research License applies. It permits non-commercial research/evaluation use; commercial use requires a separate license from the original licensor. See the full license for its conditions and NOTICE for required attribution and modification details.

Original model and documentation: Qwen/Qwen-Image-2.1-Turbo.