pipeline_tag: image-to-image
license: other
license_name: qwen-research
license_link: https://huggingface.co/rkcth/Radiance-IQ8R-from-Qwen-Image-2.1-Uncensored/blob/main/LICENSE
extra_gated_prompt: "Review the included Qwen Research License before requesting access."
extra_gated_fields:
"I accept the Qwen Research License": checkbox
base_model: Qwen/Qwen-Image-2.1
tags:
- radiance
- image-editing
- iq8r
Radiance IQ8R from Qwen Image 2.1 (Uncensored)
This draft targets the Radiance 2.1 shared-serving fork with matching ABI 41 plugins.
This community conversion uses the selected uncensored BF16 transformer, text encoder
and VAE. It packs them into a Radiance container and quantizes the image and conditioning
linears to IQ8R. The source checkpoint supplied the uncensoring changes.
The container passed checksum and metadata validation.
The 5f3bf3f3 review tree passed 1024×1024 generation and one-reference
editing on two R9700s. With Swift as the text primary, the seeded edit
matched after switching to Krea and back and after explicit unload and
cold reload.
Enable an image-size and reference limit appropriate to your GPU.
The weight payloads came from the selected public component revisiond94be5d7a09b47b0211b03d0f23d20bc18203358. The pipeline metadata, tokenizer,
scheduler and tensor layouts were taken from the Qwen release atd26bb61231c349cf6b7896fa83353113880e1ba3. The accompanying NOTICE describes
the conversion and credits the sources.
The Qwen Research License covers research and evaluation use. See the included
license for the conditions on use and redistribution.
Verify and use the download
From the downloaded directory, check the metadata and checkpoint:
sha256sum --strict -c metadata.sha256 SHA256SUMS
In the two-R9700 Swift profile, set image-model under the text model's
YAML entry to the absolute path of qwen-image-2.1-uc-iq8r.rad. The
qualified profile sets tp: 2, image-gen-max-pixels: 1048576, andimage-gen-max-refs: 1. Keep the rest of the text and memory settings
appropriate to your host. Check the YAML with radiance --config FILE --check
before starting it with radiance --config FILE.
Get the image model ID from GET /v1/models and use it withPOST /v1/images/generations or POST /v1/images/edits. Send one
reference with an edit request.
Source releases selected checkpoint components and Qwen Image 2.1.