Generate images in batch with a diffusion model

This tutorial walks you through running high-throughput batch image generation on a JSONL file of prompts using a diffusion model. You submit the text-to-image batch job bundle on GitHub to a GPU fleet on your Deadline Cloud farm. The default model is FLUX.2 Klein 4B on Hugging Face, which is Apache 2.0 licensed, ungated, distilled to 4 steps, and needs about 13 GB of video RAM (VRAM).

Each selected line in the JSONL becomes a generation task. The scheduler distributes tasks across available GPU workers, and each worker loads the diffusion pipeline once and reuses it for every task it runs. When a line carries a caption field, or a generated_text field chained from the vLLM batch inference bundle's output, the job composites the text over the generated image as crisp typography. Lines without a caption produce pure imagery, so the bundle works equally well as a plain text-to-image batch generator.

Estimated time: 30–60 minutes, including farm setup. The first run also downloads about 13 GB of model weights per worker.

Running this tutorial incurs charges for the GPU worker instances that process the job.

Overview

To complete this tutorial, follow these steps:

Set up your farm

You need a service-managed fleet with NVIDIA GPUs and at least 32 GB RAM, and a queue with a conda queue environment attached that reads CondaPackages and CondaChannels job parameters. FLUX.2 Klein 4B fits comfortably on 16 GB and larger GPUs (for example, L4 or A10G) thanks to CPU offloading.

The fastest way to get a compatible farm is to deploy the CUDA farm CloudFormation template on GitHub, which is the same template the vLLM batch inference bundle uses. After the stack reaches CREATE_COMPLETE, configure the Deadline Cloud CLI to use the new farm:

deadline config set defaults.farm_id FarmId-from-stack-outputs
deadline config set defaults.queue_id CUDAQueueId-from-stack-outputs

Prepare the input file

The input is a JSONL file with one JSON object per line. Each line must have either a prompt field or a generated_text field chained from the vLLM batch inference bundle's output:

{"prompt": "A golden hour photo of a sourdough loaf on a linen-covered table", "id": "loaf_01"}
{"prompt": "A bakery storefront at sunrise, photorealistic", "caption": "OPEN AT DAWN", "id": "shop_01"}
{"prompt": "Write a slogan for sourdough", "id": "001", "generated_text": "Real grain. Real fermentation. Real you.", "style": "rustic flat lay, warm tones"}

Optional per-line fields include caption for overlay text, style to override the job-level style suffix, font, width and height, steps, and seed. The bundle includes the same zero-dependency prompt builder tool as the vLLM batch bundle. Open tools/prompt_builder.html to build input files, import a JSONL by drag and drop, and add per-prompt overrides.

The bundled sample_prompts.jsonl file is a 10-image bakery campaign demo: bakery slogans generated by the vLLM batch inference bundle, with style hints layered on for evocative visuals.

Submit the image generation job

Alternatively, submit with the CLI:

deadline bundle submit . \
  --parameter InputFile=$PWD/sample_prompts.jsonl \
  --parameter Prompts=1-10 \
  --parameter OutputDir=$PWD/output

The Prompts and ChunkSize parameters work the same way as in the vLLM batch inference bundle, using the Deadline Cloud task chunking feature. For more information, see Task chunking for job templates. The ModelName parameter accepts anything the diffusers library can load, including SDXL Turbo and Stable Diffusion 3.5. For the full parameter list and the settings for other models, see the parameters table in the sample README on GitHub.

Download and browse the gallery

Chain from vLLM batch inference

The classic flow pairs this bundle with the vLLM batch inference bundle:

When a line has both a generated_text slogan and the original LLM prompt, the task script tries to pull the subject out of the request and use it as the visual prompt for the diffusion model, because slogans are often too abstract to give the model a concrete subject. The worker logs the visual prompt source per task so you can spot misextractions.

Clean up

To avoid ongoing charges, clean up the resources that you created for this tutorial:

Related resources

The following resources provide additional information: