ERNIE Image Turbo INT4 ConvRot
Repackaged model files for ComfyUI and others, with INT4 ConvRot converted for testing in Radiant Canvas for fast macOS inference: Radiant Canvas on the App Store.
The original model is baidu/ERNIE-Image-Turbo. The ComfyUI single-file BF16 transformer used for this conversion came from Comfy-Org/ERNIE-Image.
File
ernie-image-turbo-int4_convrot.safetensors: the diffusion transformer in ComfyUI's native ConvRot W4A4 format.
For ComfyUI, place the file in:
ComfyUI/models/diffusion_models/
Use a recent ComfyUI build with ERNIE Image and native INT4 ConvRot support. The text encoder, prompt enhancer, and VAE are not included here. Get compatible versions of those components from the original model or the Comfy-Org repack.
Conversion details
- Original model revision:
bc68c81e2a1730a394d5fc9fae70713dee940140 - Comfy-Org repack revision:
82fe29a5cd056f8b1deebc50570f125bcd4f4bea - Converter revision:
b477b06ecc3afe8c3896a6cef02f5b1b8a550a0b - Source format: BF16 single-file transformer, 409 tensors
- Quantized layers: 252 core attention and MLP weights
- ConvRot group size: 256
- Quantization group size: 64
- Output size: 4,295,963,576 bytes
- SHA-256:
b085212f916384cc2eb84d1356cecfa29106ee5b210160a093590eefabf0b9f5
The converter profile was corrected locally to preserve
adaLN_modulation.1, final_linear, and text_proj in BF16. The upstream
profile used a lowercase adaln match, while the checkpoint key uses
adaLN. Biases, normalization, embedding, input, output, and modulation
weights are also kept in BF16.
This is a test conversion. Check output quality for your own prompts and workflows before relying on it.
License
The original model is released under the Apache License 2.0. A copy of the license is included in this repository.