Why a second pass helps
Diffusion models are generally most predictable around the resolutions represented in their training. Asking an older or resolution-sensitive model to construct a much larger canvas in one pass can repeat subjects, distort anatomy, or lose composition. Hires Fix establishes the scene at a manageable size before attempting high-resolution detail.
The enlarged intermediate image then becomes the input to an image-to-image pass. This distinguishes the workflow from ordinary post-process upscaling, which can enlarge or sharpen pixels without rerunning the diffusion model over the composition.
The two stages
- Compose. Generate the initial image at the chosen base resolution using the prompt, seed, sampler, and model.
- Enlarge. Resize the latent or decoded image with the selected upscaler.
- Refine. Denoise the enlarged result for additional steps. Denoising strength determines how much the refinement may change the first pass.
Implementations expose different choices for latent versus pixel-space enlargement, second-pass steps, and target size. “Hires Fix” names the workflow, not one universal algorithm.
Where users encounter it
AUTOMATIC1111 presents Hires Fix directly in text-to-image generation. Equivalent staged workflows can be built in ComfyUI even when the interface uses different node names. It is useful when final detail matters but a direct high-resolution generation destabilizes layout or exceeds available VRAM.
Hires Fix versus upscaling
A conventional upscaler predicts higher-resolution pixels from a finished image. Hires Fix includes an upscale but then performs generative denoising, which can redraw textures, faces, and small structures. That can create genuinely new detail, but it can also change identity or introduce artifacts. A user who needs strict preservation may prefer a lower denoising value or a non-generative upscaler.
Limitations
- The second pass costs additional time and memory.
- Too much denoising can undo the first-pass composition; too little may merely enlarge defects.
- Large scale factors can expose details the model cannot reconstruct consistently.
- Newer models trained for larger native resolutions may need different settings or no equivalent fix.