Revision: 451f4fe16113bff5a5d2269ed5ad43b0592e9a14
CreativeML Open RAIL-M; review the model card. Public checkpoint; use training images you have permission to use.
Reviewed guide · no hardware test recorded
Train a LoRA
Prepare a Stable Diffusion 1.5 text-to-image LoRA training run using the pinned Diffusers training script.
Version 2026-09-27.1
train-sd15-lora · reviewed 2026-09-27 · review due 2026-10-27
Requirements
- Total download
- Unknown
- Host disk
- Unknown
- Host RAM
- Unknown
- Per-device VRAM
- Unknown
Task details and compatibility
No workload-specific hardware measurement. Dataset size, optimizer state, checkpoints and environment files add to the model download.
- model
- Stable Diffusion 1.5
- method
- LoRA
- resolution
- 512
- batchSize
- 1
- rank
- 4
- precision
- fp16
Changing task dimensions invalidates a tested-fit claim for a different task. This version has 0 recorded successful runs.
- Diffusers: 0.35.1. Pinned training script; use its example requirements.
- PyTorch, Accelerate, PEFT and Transformers: version not validated. A resolved dependency lock and GPU runtime have not been validated; record installed versions before a trial.
Artifacts and prerequisites
Revision: 0f252be0ed42006c125ef4429156cb13ae6c1d60
Apache-2.0. Public source. Install requirements.txt from the same revision.
Set up the task
On-demand GPU Pod; choose a PyTorch environment compatible with your GPU.
- Prepare and caption an authorized image dataset. Record the dataset revision and size locally; this guide does not upload it.
- Create a single-GPU on-demand Pod and verify the driver and PyTorch CUDA compatibility. Clone Diffusers and check out the exact commit listed above.
- Install Diffusers and examples/text_to_image/requirements.txt from that commit. Configure Accelerate for one GPU and fp16.
- Run the linked train_text_to_image_lora.py with the pinned base-model revision, resolution 512, train_batch_size 1, rank 4 and a persistent output_dir. Supply your dataset and explicitly choose training steps.
- Record installed package versions and test a short run. Changing resolution, batch, optimizer or dataset invalidates any future measurement for a different task.
Download, dependency installation and preprocessing time are user assumptions. Training duration and memory have not been measured.
Save outputs and shut down
- Keep your training data, model cache and output_dir on /workspace or another confirmed persistent mount.
- Allow separate capacity for checkpoints and final adapter weights.
- Wait for the final adapter save and preserve training arguments and logs.
- Download the adapter, checkpoints and any validation outputs; verify the downloaded files before shutdown.
- Stop the Pod yourself and check which storage continues to accrue charges.
NVGPU does not provision, stop or delete resources.
Plan before you launch
Limits and compatibility
- Reviewed preparation guide; dependency lock, runtime compatibility and GPU fit are pending.
- No training throughput or time-to-completion claim.
- Text-inference memory formulas do not apply to this training recipe.
Sources and review evidence
Official source: Training script, dependency installation and Accelerate setup.
Official source: Immutable script and requirements revision.