What you need before choosing a runtime
Start with the GPU you have, the task you want to complete and a supported software path. The 5.6B parameter label describes the image transformer; a usable setup also loads other components and needs memory for the output. Disk space, system RAM and GPU memory are different budgets.
Minimum versus recommended specifications
BFL identifies an INT8 ComfyUI route for NVIDIA GPUs with 8GB of VRAM. Treat that as a publisher-supported configuration to investigate, not a universal minimum for Python, every GPU or every output. A 4K job with several references can need more memory than a small text-only test.
A practical starting configuration is the current native ComfyUI template, a compatible quantized package and a small one-image workload. For a specific recommended GPU, precision or offloading strategy, use the current runtime documentation. This guide does not invent a measured VRAM table or claim that a local GPU has been tested by this site.
Choose one installation path
- ComfyUI: suitable when you want a visual graph with native model loaders. Continue with the ComfyUI workflow guide.
- Reference Python implementation: suitable when you want the publisher's CLI, Python interface or a reproducible code environment. Start from the official inference repository.
Step-by-step setup
Step 1: preserve your existing environment
Create a separate environment or installation for evaluation. Record the application version and GPU driver, and keep your working FLUX.2 workflows available. A new model should not require discarding a usable pipeline.
Step 2: confirm hardware and software support
Follow the selected runtime's installation instructions for your operating system and GPU. A CUDA-specific command is not a generic instruction for an Apple or CPU-only machine. Resolve a basic device-detection error before transferring large weights.
Step 3: download a matching package
Use the official download selector to choose the runtime, checkpoint and precision together. Accept gated repository conditions where required, and include the encoder, decoder and prompt-upsample components expected by that path.
Step 4: make the smallest useful image
Begin with a single subject and no references. Keep native settings, use a smaller output and save the seed and workflow with the result. This checks model loading, inference and file output without mixing several failure sources.
Step 5: expand the workload gradually
Test one reference edit next, then additional references or a larger image. Inspect whether preserved subjects and backgrounds remain correct. Record failures alongside successful results; a stable small run is a baseline rather than proof that every workload fits.
Use the publisher's Python CLI
The official repository documents its current Python, PyTorch and CUDA installation requirements. Use those version-specific commands from that checkout rather than combining them with packages from an unrelated tutorial. The repository also documents the weights directory and required safety and prompt-upsample models.
After following that installation and downloading its required components, the distilled text-to-image command has this shape. The prompt below is an original example; it is not a claim of a generated result.
python scripts/cli.py "A matte ceramic cup on a pale stone table, soft window light, a centered composition and a quiet neutral background" \
--model distilled --out first-image.png --seed 7Run this inside the reference implementation checkout. Keep its default prompt upsampler for an ordinary request. If a different language model prepares full scene captions, follow the publisher's documented method for that path. The official CLI examples cover editing and layout inputs; do not assume flags transfer unchanged to another runtime.
Common errors and fixes
The application cannot see the GPU
Check the runtime's device-detection output and installation requirements. Confirm you are launching the intended environment. Installing another checkpoint does not fix a mismatched driver or unsupported compute build.
Loading or sampling exceeds available memory
Separate loading failures from failures after sampling starts. Verify supported precision and documented offloading for loading problems; reduce resolution and references for a workload that starts but does not finish. Avoid changing several of those variables in the same test.
The result is generic or ignores the request
Check the prompt-upsampling path and the exact model selected. Then simplify the scene and assign each reference a single role. A running program can still have an incomplete prompting pipeline.
A previously working workflow breaks after an update
Compare application and model revisions with the saved baseline. Reopen the matching native template and reproduce the small test before restoring optional branches. Keep the result, settings and error text together when asking the runtime maintainers for help.