feat: support Qwen Image 2.1 local generation and editing - #7809
Conversation
|
Required CI is blocked by the shared server import-budget failure: 116,547 static module instantiations versus the 116,500 cap. The first failing job is https://github.com/atomantic/PortOS/actions/runs/35530318769/job/106129674884; fail-fast canceled the remaining jobs. The same failure reproduces locally, and #7807 independently reports the identical baseline count on Tracked in #7810. A bounded lazy-import experiment in the image setup module saved only seven instantiations and was reverted; the PR's image-generation implementation remains unchanged at ac06aad. Focused tests and pregate passed as documented in the PR. This PR is left OPEN, not merged over red CI. Repair the baseline under #7810, synchronize this branch, rerun the required local review if rebasing, and require the complete CI rollup before merging. |
Summary
Add Qwen-Image 2.1 to the local Image Gen catalog and Hugging Face import flow using its new
QwenImage21Pipeline. Existing registries receive the choice without replacing customized legacy entries.The renderer supports text-to-image and single-init-image editing, preserves RGBA inputs/output, passes the supported guidance parameter, and applies the new latent scaling to previews. Both runtime installers now install Transformers >=5.17 and verify the new pipeline. Documentation covers setup, the research license, and the current single-reference scope.
Upstream contract: https://github.com/QwenLM/Qwen-Image-2.1 and huggingface/diffusers#14804.
Test plan
python3 scripts/z_image_turbo_test.py: generation/editing fixture passed, including alpha preservation and native edit dispatch.bash -n scripts/setup-image-video.shandgit diff --checkpassed.npm run pregate: passed client lint and 219 selected server tests/guards.