Known limitations

← Documentation index · Settings reference · Troubleshooting

These are current boundaries, not setup failures.

Area	Current boundary	Practical path
Test Studio preparation	All seven image training families have generation workflows; each needs matching local models and native ComfyUI support	Prepare missing FLUX.1, Anima or Qwen-Image 2.1 files in Settings → Image engines → Trained image models; update ComfyUI when the node check requests it
Krea 2 img2img Studio mode	`backend/workflows/krea2_turbo_img2img.json` exists but is not wired to a separate Studio mode	Krea 2 Edit remains available for dataset/reference generation; Studio uses its reachable text-to-image path
Canvas comparison generation	A single launch requires checkpoints from one model family	Run separate same-family comparisons; cross-family workflows do not share one base graph
Dual captions	Local training only; Krea 2 and Anima cache text embeddings and use the long caption alone	Keep the long caption complete for every family
Concept datasets on SDXL	Concept captions are only produced as prose; there is no booru-tag concept captioner, and the prose/booru selector is hidden on concept datasets. A concept dataset trained on a booru-native SDXL checkpoint therefore always trips the caption-style check	Train the concept on a prose family (Z-Image, Krea 2, FLUX.1, FLUX.2 Klein, Anima), or force the launch and accept the quality loss a booru-native base takes from prose. Character and Style datasets are unaffected — both reach booru captions normally
Cloud training	Support is family- and variant-specific, and full-state continuation is local-only	Trust the launch UI's live compatibility reason before renting a pod
GPU Docker	Ollama and local ai-toolkit training are not bundled; watermark inpainting is not currently available in this lane	Connect external Ollama/ai-toolkit, use cloud training, and use model-free watermark crop where suitable
Remote access	The app has no user accounts and ComfyUI's published port is not protected by the LDS access token	Keep loopback defaults or use tokens, firewall/VPN and an authenticated reverse proxy as described in Security
Browser-local preferences	Some convenience choices, such as the last generator, are remembered per browser	Explicitly select a configured engine after moving to a new browser/profile

ComfyUI-dependent paths are covered extensively against a mocked API, but not every model/custom-node combination has been exercised on live third-party installations. A failed preflight should name the missing asset; use Settings → Local tools → Test and attach the diagnostic report when a supported layout is not detected.

Provider policies, moderation and service availability are outside this project's control. See the direct Gemini, ChatGPT subscription, OpenRouter/image-engine and Pexels notes before depending on those lanes.