Troubleshooting

Starting up

“is not recognized as an internal command”

Make sure the whole archive was unzipped: CoCo Image Studio.exe should sit next to the data folder, which holds CoCo.bat, the core, server, web and tools folders and requirements-lock.txt.

“cannot see your GPU”

The PyTorch version is right but it can’t see the graphics card, so reinstalling won’t help. It’s the driver or the hardware:

1. Update your NVIDIA driver to 580 or newer.
2. Check that the card looks normal in Device Manager.

The library still works in this state; only generating fails.

RTX 50 series: “this PyTorch doesn’t support your GPU (sm_120)”

Delete data\venv\.coco-ready and open CoCo again. It checks the installed version and reinstalls the right one if it doesn’t match.

Generating

Out of graphics memory, or very slow

Prerequisite: enough system memory

The “Balanced” and “Low” VRAM modes, and the automatic streaming used when a large model doesn’t fit on the card, keep part of the model in system memory and move it to the graphics card only when needed. If system memory also runs out, Windows falls back to virtual memory on disk and generation becomes much slower. While generating, check usage in Task Manager under “Performance → Memory”. If it is nearly full, close other programs first, or turn on “Store the main model in fp8” from step 2 (it also reduces system memory use). 32 GB of system memory is recommended for large models such as Z-Image, Krea 2, Qwen Image and FLUX.

Try these in order:

1. In the generate panel, set More → VRAM to “Balanced”.
2. Turn on “Store the main model in fp8” in ⚙ Settings → Generate (the main model takes about half the memory; images change slightly).
3. Set More → VAE tiled decode to “Always”.
4. If it’s still not enough, set VRAM to “Low” (uses the least, but is several times slower).

When a large model (Z-Image, Krea 2, Qwen Image, FLUX) doesn’t fit on the card, CoCo streams it automatically instead of failing — it’s just slower.

A LoRA has no effect or fails to load

• Check the “Messages” tab in the generate panel; the reason is there.
• LoRAs for a different architecture than the main model (say an SDXL LoRA on Anima) are skipped.
• LyCORIS (LoHa, LoKr) isn’t supported yet.

V-Pred models (NoobAI V-Pred and similar)

Just pick it. CoCo recognises it from the model file and switches to the right method. The official recommendation is Euler, CFG 4–5, 28–35 steps; if you pick a sampler that doesn’t suit it, CoCo switches to Euler and says so in the messages.

Library

Images show “No generation settings”

Their metadata was stripped (usually downloaded or edited images). That’s not an error.

New images don’t show up

Images made in other tools appear the next time you start CoCo, or after “⋯ → Rescan” in the library. Also check that their folder is listed under ⚙ Settings → Library → Library folders.

Still stuck

Report it in the GitHub Issues (the link will be added when the public version is released). Including what the “Messages” tab says makes it much quicker to find the cause.