10 The Flux Workflow

This chapter introduces Flux as a powerful, open image generation model and walks through how to use it effectively in ComfyUI. It contrasts Flux’s workflow with familiar Stable Diffusion pipelines, highlighting its modular model loading, new guidance mechanism, and a more nuanced sampling stack. The focus is practical: get set up, understand what’s different, and learn the few levers that most strongly shape output quality, style adherence, and realism.

Setup centers on loading separate components rather than a single checkpoint: two CLIP encoders (for stronger prompt understanding), a VAE, and the Flux.1-dev UNet. Because Flux is memory-intensive, the chapter recommends an fp8 quantized UNet variant to substantially cut VRAM usage with only minor quality trade-offs, making high-end consumer GPUs and unified-memory Macs more viable. It also explains the DualCLIPLoader’s role in prompt adherence, notes when custom VAEs matter, and underscores that you can run the workflow without mastering every node—once the files are in place, you can prompt and go.

Flux’s “Guidance” replaces traditional CFG behavior and has a larger, more direct impact: lowering it can both improve style fidelity (e.g., manuscript-like outputs) and reduce the “plasticy” look in photorealistic renders. Sampling is more involved: RandomNoise adds seed control modes (fixed, increment, decrement, randomize); ModelSamplingFlux introduces max_shift (highly impactful on color separation and overall look) and base_shift (subtle and sometimes no-op at certain resolutions); and BasicScheduler exposes denoise, which changes results dramatically within a narrow range and should be adjusted in small steps. The chapter closes by noting batch-size constraints, resolution inputs, and the overarching theme: experiment with Guidance, max_shift, and denoise to steer Flux, while using quantization to fit your hardware.

Drag this image to ComfyUI to get the Flux workflow.

The Flux workflow is notably different than SD1.5 and SDXL

Our adorable first image from Flux

The Load Diffusion Model node allows you to select the U-NET.

Using the fp8 model only minorly impacts the quality of our output.

The DualCLIPLoader allows us to use two CLIP encoders.

FluxGuidance and BasicGuider nodes.

The Guidance parameter in Flux can help the model adhere to style better.

Lower Guidance also helps to make images look less “plasticy”.

Sampling is notably more complex when working with Flux.

RandomNoize adds a control_after_generation option.

Noise Schedule when max_shift and base_shift both are 0.0.

With max_shift of 1.5, the noise removal is no longer uniform.

Max shift can have a notable impact on the resulting image generated.

Base shift slightly modifies the denoising curve.

While there is a difference in images, it is more subtle.

Under some aspect ratios, base_shift has no impact at all!

Denoise should be changed in very small increments

• The Flux workflow has quite a few differences from the standard Stable Diffusion workflow.
• When working with models requiring high memory, it is helpful to look for quantized versions of these models.
• Part of the reason Flux is so good at prompt adherence is because it uses two CLIP encoders.
• Guidance in Flux is similar to CFG in SD1.5 and SDXL, but doesn’t allow for a negative prompt.
• Reducing Guidance can help the model pay more attention to the style recommended in the prompt.
• Flux allows for much more nuanced control over scheduling.
• Max shift can have a pretty major impact on your final result, but base shift is less impactful and in many cases will have no impact at all.
• Reducing the Denoise value can make the image look nicer, but should only be done in very small steps as it can quickly degrade the image.

FAQ