FireRed Image Edit LoRA Guide

The official FireRed Image Edit LoRA collection extends FireRed-Image-Edit-1.1 for three distinct goals. Pick the adapter by task instead of stacking every file into one workflow.

Official LoRA Zoo

The FireRed Image Edit LoRA Zoo currently describes three adapter families:

• CoverCraft targets artistic typography, decorative lettering, cover design, and poster-style title layouts.
• Makeup targets cosmetic edits such as skin refinement, color harmony, and eye or lip detail while preserving identity.
• Lightning (8-step distilled) targets faster inference with an eight-step workflow while retaining instruction alignment and visual coherence.

CoverCraft and Makeup change task behavior. Lightning changes the inference path. Evaluate them separately before combining adapters.

Install an adapter in ComfyUI

1. Verify that the adapter was trained for FireRed-Image-Edit-1.1 or for the exact base revision named by its official model card.
2. Place the .safetensors adapter in ComfyUI/models/loras.
3. Refresh ComfyUI's model list or restart the application.
4. Add the LoRA through the model-only LoRA loader used by the official workflow.
5. Begin with the adapter strength and sampler settings supplied by the matching workflow, then change one value at a time.

Always run the base workflow first. If the base model already fails to follow the prompt, adding a LoRA makes diagnosis harder.

CoverCraft prompts

Describe the exact text, hierarchy, medium, and layout. For example:

Replace the title with “MIDNIGHT BLOOM”, large hand-painted serif lettering, centered magazine cover layout, preserve the portrait and background lighting.

Check spelling at full resolution. Typography adapters can improve style and structure, but generated text still needs visual review before commercial use.

Makeup prompts

Protect identity explicitly and specify only the cosmetic changes you need:

Keep facial identity, expression, hairstyle, and skin texture. Apply soft editorial makeup with warm blush, subtle eyeliner, and natural rose lips.

Use a clear, front-facing source for your first test. Large pose changes, aggressive retouching, and multiple simultaneous style requests make identity drift harder to isolate.

Lightning eight-step inference

Use Lightning when throughput matters and the official matching workflow is available. Enabling the LoRA without changing the sampling configuration does not reproduce the intended distilled path. Conversely, running only eight steps without the correct Lightning adapter can reduce edit quality.

Record the transformer revision, LoRA filename, strength, sampler, scheduler, CFG, step count, and seed with every production preset. That makes regressions traceable when model files or ComfyUI nodes change.

Compatibility checklist

• Base transformer and LoRA target the same FireRed revision.
• The workflow uses the text encoder and VAE named by the official package.
• Only one new adapter is introduced per A/B test.
• The prompt, seed, steps, and resolution stay fixed during comparison.
• Output is checked for identity, instruction following, text accuracy, and unintended region changes.

For the model's base capabilities and Apache-2.0 license, see the official FireRed-Image-Edit-1.1 page.