File size: 3,644 Bytes
55f86a4 39d9f01 f084c4c ed034c2 f084c4c 2f61a94 7ebfb23 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 | ---
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 |
|