| --- |
| license: apache-2.0 |
| base_model: |
| - OpenMOSS-Team/MOSS-Image-Base |
| - OpenMOSS-Team/MOSS-Image-SFT |
| library_name: diffusers |
| tags: |
| - simpletuner |
| - moss-image |
| - lora |
| - experiments |
| --- |
| # MOSS-Image: A LoRA Experiment Journal |
|
|
| This experiment compares learning Domokun with MOSS-Image Base and SFT at 512px and 1024px, |
| then repeats the 512px comparison with vision-grounded native JSON captions. |
|
|
| ## Protocol |
|
|
| The initial four runs use the original filename captions for all 29 images in |
| [RareConcepts/Domokun](https://huggingface.co/datasets/RareConcepts/Domokun), pinned to |
| `2e4831a72f2a701c080e0046c8f8ec41cb1ccbcf`. Each run starts from its released checkpoint: |
| Base `1ceede80badddd6a7812069619edaa5633587192`, SFT `fc5de0d56fc910cbca6f6625a70f5f582b26deb1`. |
| Resolution is equivalent image area with original aspect ratios preserved. |
|
|
| Every run targets 10,000 optimizer updates on one H100 80 GB, with rank-32 BF16 LoRA, |
| batch 4, gradient accumulation 1, learning rate 0.0001, 25 warmup updates, seed 42, and REPA. |
| Validation uses four prompts, 50 sampling steps, CFG 3, seed 42, at startup and every 500 updates. |
| All 20 saved adapter checkpoints and their validation previews are published per completed run. |
| Baseline images are available in each run repository's `baseline` folder. |
|
|
| The two JSON-caption runs are queued after all initial runs finish and start fresh from Base/SFT. |
| They share one set of captions generated using Qwen/Qwen3.5-27B, revision |
| `fc05daec18b0a78c049392ed2e771dde82bdf654`, grounded in training images and actual 512px bucket dimensions. |
| Captions must pass native schema, bounding-box, concept-token, and tokenizer-length checks before training. |
| These runs use the same 10,000-update protocol; no 1024px JSON runs or assistant-LoRA runs are scheduled. |
|
|
| Adapters and previews are collected without interrupting training. Full training state and logs are archived separately. |
| The public recipes retain the in-process publishing flags used during training; an external collector uploads results. |
| The original-caption runs use SimpleTuner revision `e30354dffd02c80f52f39912cea271f12be5ac4e`; |
| the JSON runs use the shared schema implementation from [PR #3273](https://github.com/bghira/SimpleTuner/pull/3273). |
|
|
| ## Progress |
|
|
| | Run | Observed updates | Latest published checkpoint | Status | |
| | --- | ---: | ---: | --- | |
| | [base-512](https://huggingface.co/SimpleTuner/moss-image-base-domokun-512) | 10,000 / 10,000 | [10,000](https://huggingface.co/SimpleTuner/moss-image-base-domokun-512/tree/main/checkpoint-10000) | completed | |
| | [base-1024](https://huggingface.co/SimpleTuner/moss-image-base-domokun-1024) | 10,000 / 10,000 | [10,000](https://huggingface.co/SimpleTuner/moss-image-base-domokun-1024/tree/main/checkpoint-10000) | completed | |
| | [sft-512](https://huggingface.co/SimpleTuner/moss-image-sft-domokun-512) | 10,000 / 10,000 | [10,000](https://huggingface.co/SimpleTuner/moss-image-sft-domokun-512/tree/main/checkpoint-10000) | completed | |
| | [sft-1024](https://huggingface.co/SimpleTuner/moss-image-sft-domokun-1024) | 10,000 / 10,000 | [10,000](https://huggingface.co/SimpleTuner/moss-image-sft-domokun-1024/tree/main/checkpoint-10000) | completed | |
| | [json-base-512](https://huggingface.co/SimpleTuner/moss-image-base-domokun-json-512) | 10,000 / 10,000 | [10,000](https://huggingface.co/SimpleTuner/moss-image-base-domokun-json-512/tree/main/checkpoint-10000) | completed | |
| | [json-sft-512](https://huggingface.co/SimpleTuner/moss-image-sft-domokun-json-512) | 10,000 / 10,000 | [10,000](https://huggingface.co/SimpleTuner/moss-image-sft-domokun-json-512/tree/main/checkpoint-10000) | completed | |
|
|
