CoolFace
Modelpublic

ApacheOne/orbitquant_unets

sourceHugging Faceotherupdated 2mo agoView on Hugging Face
0likes
Model Card

USE CONVROT INT4 and INT8 for now

OrbitQuant SDXL Exports

Experimental OrbitQuant and ConvRot quantized exports for CreaPrompt Lightning Hyper-SDXL.

Work in progress: The native A4 runtime is not public yet. These models cannot currently be used without the custom runtime. A ComfyUI loader node that outputs a model compatible with the core KSampler is still under development.

W2A4 Recovery Status

WIP for W2A4 block-level rollout recovery.

Model Format

These are not standard SDXL checkpoints that can be loaded directly by existing SDXL loaders.

Depending on the export, the files can contain:

  • —OrbitQuant-packed SDXL UNet weights
  • —ConvRot-quantized SDXL UNet weights
  • —A4 activation-quantization metadata
  • —Quantized CLIP and text-encoder weights
  • —FP16 VAE weights
  • —__orbitquant__ metadata
  • —Codebooks
  • —Rotations
  • —Quantization scales
  • —Runtime configuration

A custom runtime is required to execute the packed OrbitQuant and ConvRot weights without expanding them back into ordinary FP16 weights.

Examples

Full OrbitQuant versus Full ConvRot

<table> <tr> <td align="center"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/comparisons/FULLOQW4A4vsFULLCRW4A4CLIPINT4comparison.png" width="760"

<br> <strong>Full OrbitQuant W4A4 versus Full ConvRot W4A4 — CLIP INT4</strong> </td> </tr> </table>

Full-UNet Tests

<table> <tr> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/fullunet/convrotw4a4clipint4/FULLUNETCRW4A4CLIPINT4_test.png" width="420"

<br> <strong>Full ConvRot W4A4 UNet — CLIP INT4</strong> </td> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/fullunet/orbitquantw4a4clipint4/FULLUNETOQW4A4CLIPINT4_test.png" width="420"

<br> <strong>Full OrbitQuant W4A4 UNet — CLIP INT4</strong> </td> </tr> </table>

Mixed OrbitQuant and ConvRot UNet — CLIP FP16

<table> <tr> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/mixedunet/clipfp16/creapromptSDXLOQW4A4CRW4A4CLIPFP16_test.png" width="420"

<br> <strong>OrbitQuant W4A4 + ConvRot W4A4 — CLIP FP16</strong> </td> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/mixedunet/clipfp16/creapromptSDXLOQW4A4CRW8A8CLIPFP16_test.png" width="420"

<br> <strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP FP16</strong> </td> </tr> </table>

Mixed OrbitQuant and ConvRot UNet — CLIP INT4

<table> <tr> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/mixedunet/clipint4/creapromptSDXLOQW4A4CRW4A4CLIPINT4_test.png" width="420"

<br> <strong>OrbitQuant W4A4 + ConvRot W4A4 — CLIP INT4</strong> </td> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/mixedunet/clipint4/creapromptSDXLOQW4A4CRW8A8CLIPINT4_test.png" width="420"

<br> <strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP INT4</strong> </td> </tr> </table>

Mixed OrbitQuant and ConvRot UNet — CLIP INT8

<table> <tr> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/mixedunet/clipint8/creapromptOQW4A4CRW8A8_test.png" width="420"

<br> <strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP INT8</strong> </td> <td align="center" width="50%"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/mixedunet/clipint8/firstmodeltest.png" width="420"

<br> <strong>First Mixed-Quant Model Test — CLIP INT8</strong> </td> </tr> </table>

Root Example

<table> <tr> <td align="center"> <img src="https://huggingface.co/ApacheOne/orbitquantunets/resolve/main/hybridquants/orbitW4A4convrotW4A4quant/examples/creapromptSDXLOQW4A4CRW8A8CLIPINT4_test.png" width="520"

<br> <strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP INT4</strong> </td> </tr> </table>

Runtime Availability

The following components are still required for normal use:

  • —Native packed OrbitQuant weight execution
  • —Native A4 activation quantization
  • —Native ConvRot quantized operators
  • —ComfyUI model loader
  • —Integration with the core ComfyUI KSampler
  • —Runtime validation
  • —Performance tuning

Until the custom runtime is released, these files should be considered experimental development exports.