fix(qwen): fix ghosting artifacts Qwen Image Edit

[lstein]

Summary

• Qwen Image Edit was producing a ghost/doubling artifact across the whole frame outside the masked edit region. Root cause: qwen_image_denoise.py bilinear-resized the reference latents to the noisy latent's dimensions and used identical img_shapes tuples for both segments. QwenEmbedRope varies its spatial RoPE frequencies by each segment's H/W, so identical dims gave both segments the same spatial positions and cross-attention couldn't disentangle them.
• The denoise now keeps reference latents at their own (H, W), packs them at those dims, and places (1, ref_h // 2, ref_w // 2) in the reference segment of img_shapes — matching diffusers' QwenImageEditPipeline (pipeline_qwenimage_edit.py:755-760) and QwenImageEditPlusPipeline (pipeline_qwenimage_edit_plus.py:743-751).
• The reference qwen_image_i2l is now resized to ~1024² area preserving aspect ratio (matching diffusers' VAE_IMAGE_SIZE) so the reference token sequence stays in the distribution the model was trained on.
• Defensive: qwen_image_image_to_latents.py bumps multiple_of from 8 to 16 so VAE-encoded latents always have even spatial dims (required by the 2×2 patch packing); a ValueError is raised if a directly-wired reference latent has spatial dims that align to less than 2.

Test plan

• Backend: uv run pytest tests/app/invocations/test_qwen_image_denoise.py — 15 pass (8 new: _align_ref_latent_dims validation incl. zero-dim guard, _build_img_shapes distinct-dims regression)
• Frontend: pnpm test:no-watch buildQwenImageGraph — 30 pass (6 new: calculateQwenImageEditRefDimensions matches diffusers calculate_dimensions for square/landscape/portrait/extreme-ratio inputs; verifies computed dims land on the ref i2l node)
• Frontend lint: pnpm lint:eslint, pnpm lint:prettier, pnpm lint:knip, pnpm lint:tsc — clean
• Backend lint: uv run ruff check / uv run ruff format --check on changed files — clean
• Broader Qwen tests pass: text encoder, model loader, LoRA conversion utils, main config, GGUF variant detection, double variant regression
• Manual: Using the full Qwen Image Edit diffusers model, upload a reference image and prompt for a targeted change, "e.g. change the subject's tee-shirt from blue to red.".
• Manual: same with one of the quantized (GGUF) transformers, with and without applying a Lightning LoRA — confirm artifact is gone in both.
• Manual: regular Qwen Image txt2img / img2img / inpaint / outpaint flows still produce expected output (the multiple_of=16 i2l change is a no-op for canvas-composited paths but worth a smoke test).
• Manual: try with a reference image whose aspect ratio differs significantly from the output dimensions. It should renderwithout distortion.

🤖 Generated with Claude Code

[JPPhoto]

One code issue to be resolved prior to functional testing:

• invokeai/app/invocations/qwen_image_denoise.py:372-395, invokeai/app/invocations/qwen_image_image_to_latents.py:36-43,83-89, invokeai/frontend/web/src/features/nodes/util/graph/generation/buildQwenImageGraph.ts:198-210: the fix only clamps the reference image to the diffusers-style ~1024^2 size when the graph is built by the updated frontend. The backend qwen_image_denoise path now preserves and packs whatever reference latent size it is given, but qwen_image_i2l still defaults to encoding at the image's original size unless width and height are explicitly set. Direct API/manual graph callers, and any existing graph JSON that does not populate those fields, will still send native-resolution reference latents into the transformer. In those cases the model still receives out-of-distribution reference sequence lengths, so the original artifact can persist and VRAM usage can spike on large reference images. To expose this issue, add a backend integration test that wires qwen_image_i2l to qwen_image_denoise without frontend-provided width/height and asserts the ref segment is resized/clamped to the diffusers-derived dimensions instead of staying at native size.

[lstein]

Good catch, addressed in eaf116e — added _maybe_clamp_ref_latent_size in qwen_image_denoise.py that bilinear-downscales the reference latent to diffusers' calculate_dimensions(1024², aspect_ratio) (snapped to multiples of 32 in pixel space) whenever the input exceeds the budget. It runs before _align_ref_latent_dims, so direct API/manual graph callers, older graph JSON without width/height on qwen_image_i2l, or any other producer of native-resolution reference latents all end up at the same dims the frontend was already producing — and the same dims diffusers feeds the transformer. In-budget latents (≤ 128×128 = 1024² pixels) pass through untouched.

7 new unit tests in TestMaybeClampRefLatentSize cover the matrix you asked about: in-budget (unchanged), native landscape (1600×1200 → clamped to 1184×896 → 148×112 latent), native portrait, huge (4096² → 1024²), extreme aspect ratio (1920×1080 → 1376×768), and a parametric check that every clamp output is packable (multiple of 4 in latent space, so always even after _align_ref_latent_dims).

I didn't add a full I2L→denoise integration test because the i2l step requires loading the Qwen VAE model — heavy for CI. The unit tests assert the exact same invariant the integration test would: given a too-large reference latent, the dims fed to packing/img_shapes are the diffusers-derived ones, not native.

[JPPhoto]

I was able to do testing with the quantized edit model with and without the Lightning LoRA and it worked with a different source image aspect ratio than output. I think this change is good for the next RC and public testing!