[Bug] Ideogram 4: output collapses to a horizontal center band at 1024×1024 (correct at 1536²/2048²) #1648

Activity

1. Meeseeks-Mr

   added

   bugSomething isn't workingSomething isn't working

   on Jun 13, 2026
2. stduhpf commented on Jun 15, 2026

   

   stduhpf

   on Jun 15, 2026

   Last edited by stduhpf

   Contributor

   More actions

   Is this a fully auto-genereated issue? Not only is the markdown formatting completely broken but I feel like no one (not even a VLM) actually bothered to look qualitatively at the generated images. If you have an automated system to detect and report issues, at least use AI to format the text and to actually look at the results.

   Ideogram4 will often generate that kind of image if the prompt is not formatted exactly as the model expects. In this case, it's missing the "compositional_deconstruction" field. If the same prompt is working at 1536x1536 it's just random luck, some other seed values might also generate the image correctly at 1024x1024.

   For context, here's the 1024x1024 image generated with the same prompt and seed 42 (using Q6_K model instead of Q4_K):

   

   Here is it with the same settings, but using a tweaked prompt that follows the official prompting format:

   prompt: '{"high_level_description":"A studio product photograph of a single fluffy orange cat next to a small wooden sign that reads ideogram4,","style_description":{"aesthetics":"","lighting":"soft diffused studio light","photo":"","medium":"photograph"},"compositional_deconstruction":{"background":" soft daylight, clean light-gray background","elements":[{"type":"obj","bbox":[80,80,920,600],"desc":"A fluffy orange cat"},{"type":"obj","bbox":[540,370,950,990],"desc":"a wooden sign"},{"type":"text","bbox":[580,430,890,930],"text":"Ideogram4","desc":"black paint on the wooden sign"}]}}'
   

   The original issue comment with fixed formatting

   Git commit

   The commit I am currently using is 9b0fceb. This corresponds to the master-691-563137a-1-g9b0fceb version of the sd.cpp project.

   Operating System & Version

   sd.cpp version	master-691-563137a-1-g9b0fceb (commit 9b0fceb)
   Backend	CUDA (-DSD_CUDA=ON -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89)
   GPU	NVIDIA GeForce RTX 4080 Laptop (12 GB), compute capability 8.9
   CUDA toolkit	12.8
   OS / compiler	Windows 11, MSVC 19.29 (VS 2019 Build Tools), Ninja

   GGML backends

   CUDA

   Command-line arguments used

   ./sd-cli --diffusion-model ideogram4-Q4_K.gguf --uncond-diffusion-model ideogram4_uncond-iQ4_NL.gguf --llm Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae flux2_ae.safetensors -p '{"high_level_description":"A studio product photograph of a single fluffy orange cat next to a small wooden sign that reads ideogram4, soft daylight, clean light-gray background", "style_description":{"lighting":"soft diffused studio light"}}' -W 1024 -H 1024 --steps 20 --cfg-scale 7.0 --diffusion-fa -v -o out_1024.png

   Steps to reproduce

   ./sd-cli \
     --diffusion-model ideogram4-Q4_K.gguf \
     --uncond-diffusion-model ideogram4_uncond-iQ4_NL.gguf \
     --llm Qwen3VL-8B-Instruct-Q4_K_M.gguf \
     --vae flux2_ae.safetensors \
     -p '{"high_level_description":"A studio product photograph of a single fluffy orange cat next to a small wooden sign that reads ideogram4, soft daylight, clean light-gray background", "style_description":{"lighting":"soft diffused studio light"}}' \
     -W 1024 -H 1024 --steps 20 --cfg-scale 7.0 --diffusion-fa -v -o out_1024.png
   
   # Same command with -W 1536 -H 1536  -> correct image
   # Same command with -W 2048 -H 2048  -> correct image

   What you expected to happen

   a full-frame image at 1024×1024 like the one produced at 1536/2048.

   What actually happened

   Actual at 1024×1024:** content only in a horizontal strip at the vertical center; the rest is a uniform fill.

   Logs / error messages / stack trace

   Quantitative evidence — luminance standard deviation per cell on an 8×8 grid
   (0 = perfectly flat). 1024² shows signal only in the two center rows:

    1024×1024 (BROKEN)                  1536×1536 (CORRECT)|
      1  1  1  1  1  1  1  1              5  5 23 48 47 24  5  5
      0  0  0  0  0  0  0  0              6  6  4  4  4  4  5  5
      0  0  0  0  0  0  0  0              5  5  4 69 74  4  4  5
     33 50 49 51 45 53 54 34             4  4  4 72 73  4  4  3 
     28 54 46 48 46 47 46 30             3  2 26 55 54 34  3  4 
      0  0  0  0  0  0  0  0              3  3 22 40 39 30  6  4 
      0  0  0  0  0  0  0  0              4 12 39 61 58 49 10  4 
      1  0  1  1  1  0  1  0              4  4 34 63 46 53  9  5 
    overall stddev ≈ 23 (banded)         overall stddev ≈ 43 (distributed) 
   

   Additional context / environment details

   What I ruled out (the band is invariant to all of these)

   The center-band collapse at 1024² is identical across every variable below; only
   changing the resolution fixes it:

   • Quantization: Q6_K+Q2_K and Q4_K+iQ4_NL — identical.
   • CFG scale: 1.0 (gray), 3.5, 7.0 — band present at all (1.0 is fully gray).
   • --guidance: present (3.5) vs. omitted — identical.
   • Flash attention: --diffusion-fa on vs. off — identical (off is lower contrast).
   • --flow-shift: auto vs. forced 3.0 vs. forced 6.0 — band persists.
   • Prompt: sparse one-liner vs. rich structured JSON — band persists (rich prompt
     changes the band's content but not the collapse).
   • Offload: --offload-to-cpu on vs. off — identical.

   Resolution sweep (same command, varying -W/-H)

   Resolution	Result	overall stddev
   1024²	broken (center band)	~12–23
   1280²	partial (some spread, still weak)	~33
   1536²	correct	~44
   2048²	correct	~43

   Notes

   • Sampling runs cleanly (20 Euler steps, no NaN warnings), the VAE decode completes,
     and the run exits 0 — the output is simply spatially collapsed on the height axis
     at low resolution.
   • Ideogram 4 support is recent (added in feat: add ideogram4 support #1609, then refactored in refactor: unify model config detection #1613 / refactor: reorganize src model layout #1615);
     this build (9b0fceb) post-dates those. Happy to test patches or provide the full
     verbose -v logs and sample PNGs.