[Bug] FLUX.1-Kontext-dev: -r reference editing outputs pure noise (t2i works fine) on Vulkan and CPU, both Q4_K_M and Q8_0 GGUF #2133

Summary

FLUX.1-Kontext-dev reference-image editing (-r) produces pure RGB noise with every parameter/backend/quantization combination I tried, while text-to-image with the same model files works perfectly. Reproduced on both the latest release and the build before #2085.

Environment

• stable-diffusion.cpp: master-959-d637cfb (win-vulkan-x64 release), also reproduced on master-955-b7a74e6
• OS: Windows 11, AMD RX 6750 GRE 12GB (RDNA2) via Vulkan; also reproduced with --backend cpu (pure CPU)
• Models: flux1-kontext-dev-Q4_K_M.gguf and flux1-kontext-dev-Q8_0.gguf from QuantStack/FLUX.1-Kontext-dev-GGUF (byte sizes identical to upstream), ae.safetensors (verified 16-ch latent Flux VAE), clip_l.safetensors, t5-v1_1-xxl-encoder-Q8_0.gguf

What works

Text-to-image, same model, same session — clean output in 37s:

sd-cli.exe --diffusion-model flux1-kontext-dev-Q8_0.gguf --vae ae.safetensors --clip_l clip_l.safetensors --t5xxl t5-v1_1-xxl-encoder-Q8_0.gguf -p "a lovely cat holding a sign says 'hello'" --cfg-scale 1.0 --guidance 2.5 --steps 12 --sampling-method euler -W 512 -H 512 -o ok.png

What fails

Reference editing (official example style command) — output is pure RGB noise:

sd-cli.exe -r ref.png --diffusion-model flux1-kontext-dev-Q8_0.gguf --vae ae.safetensors --clip_l clip_l.safetensors --t5xxl t5-v1_1-xxl-encoder-Q8_0.gguf -p "Change the sign text to 'KONTEXT OK'" --cfg-scale 1.0 --sampling-method euler --clip-on-cpu -o noise.png

Already ruled out

• --cfg-scale 1.0 (official recommendation) vs 2.5; --guidance default vs 2.5
• scheduler: default vs simple; steps 8 / 12 / 20 / 24
• resolution: 256x256, 512x512 (matching 512x512 ref), and default (ref native 1380x900 — that run produced a blank/white image instead of noise)
• backend: Vulkan and --backend cpu
• quantization: Q4_K_M and Q8_0
• sd.cpp version: master-955 (before feat: run conditional and unconditional CFG in one batched UNet forward #2085 CFG batched forward) and master-959 d637cfb
• VAE file integrity verified (safetensors header, decoder.conv_in.weight [512,16,3,3], 245 keys); Q8_0 gguf byte-size matches upstream exactly
• sd-server img2img API with Qwen-Image-Edit on the same install works fine (its ref/img-cond path is OK), so the issue appears specific to the Kontext ref path

Expected

Edited image preserving the reference, per docs/kontext.md example.

Happy to provide more logs/screenshots if needed.