Project
requiredA project is one folder holding the downloaded materials, the corpus, the precomputed shards, checkpoints and the exported GGUF. Everything resumes from that folder after a reboot.
Materials
teacher · student · datasetsCorpus
text promptsImage datasets
downloaded as materials; instructions are synthesized from the captionsExtra image folders
jpg/png files, optional <name>.txt caption next to each imageText-only prompts
for the text-to-image template share; the same public prompt datasets / filesExtra local prompt files
one prompt per line, or .jsonl with a "text" fieldUpsampled captions
a share of the prompts rewritten by FLUX.3's prompt-upsampler recipe on a running gguf-server (Qwen3.5 GGUF; --parallel N --ctx-size N×12288, the system prompt is ~10k tokens), each caption one more sample; cached in data/captions.jsonlPrecompute
teacher + student▾
Training
adapter▾
Steps is not locked in: change it here (Save settings) or on the Train tab at any point. While training runs the new count takes effect within a log interval — lower it to finish (validate, save, export) at the current step, raise it to keep going; after a finished run a higher count resumes from last.pt. The cosine lr schedule is computed from the count in force, so extending a finished run lifts the lr back up mid-schedule.
Output
f16 GGUFPipeline
no projectOnce training has produced best.pt, the Export GGUF and Evaluate boxes are clickable: they regenerate the GGUF / eval.json (kept in the project folder) from the checkpoint at any time, e.g. after the file was deleted.
You do not have to wait for the last step: Stop pauses the run after the current step (Start / Resume continues it), and Export snapshot writes a usable <name>-step<N>-f16.gguf from the current step at any moment — also while training runs. With Keep steps for later export on (Setup › Train, off by default to save disk) the trainer also keeps weights-only checkpoints on a step schedule of your choice (checkpoints/history/), so after the run any kept step — not only the one you paused at — can be exported from the Output tab and compared. Snapshots are listed there with their validation score, so the steps-vs-quality trade-off is visible instead of hidden behind the final export.
Training metrics
Live hardware
Hardware
Environment
Pipeline log
Exported files
Evaluation
Kept steps
Weights-only checkpoints the trainer keeps when Keep steps for later export is on (Setup › Train: every N steps inside the chosen step window, plus stops, snapshot requests and the final step; checkpoints/history/step<N>.pt). Export writes that step as a step-tagged GGUF — exactly the file a snapshot taken at that moment would have produced — so any completed point of a run can be extracted and compared later, not only the one it was paused at. Resetting training deletes them.
Snapshots
Mid-run exports of the same adapter, one per step: try any of them in the engine with the command below (swap the file name) and see how much of the final quality is already there after a fraction of the steps. val cos is the validation cosine measured when the checkpoint was saved; eval is the full evaluation of the exported GGUF when it was requested. Training can continue or resume after every snapshot.