FireRed Image Edit FP8 vs GGUF vs BF16

Compare FireRed Image Edit 1.1 FP8, GGUF, and BF16 files by source, memory use, compatibility, and quality tradeoffs before downloading.

Choosing FireRed Image Edit FP8 vs GGUF is mainly a compatibility and memory decision. Start from the official BF16 release as the quality reference, then use a quantized build only when your hardware or workflow needs it.

Quick comparison

FormatBest fitMain advantageMain caution
BF16High-memory GPU and reference testingOfficial baseline with the least quantization lossLargest download and memory requirement
FP8Supported GPU/runtime with a trusted conversionLower transformer memory than BF16FP8 builds may be community conversions and runtime support varies
GGUFComfyUI-GGUF or another compatible quantized loaderMultiple quantization levels let you trade memory for fidelityRequires the correct loader and can introduce stronger quantization loss

The official FireRed-Image-Edit-1.1 model card identifies the base model as a 20B-parameter BF16 release. The official ComfyUI repository publishes the workflow and lists BF16 plus quantized GGUF options. Treat any FP8 file as a separate conversion unless the FireRed Team source explicitly identifies it as official.

Official file-size table

The following sizes come from the official FireRed Team ComfyUI repository metadata. They are download sizes, not promised peak-VRAM figures: ComfyUI also loads the text encoder, VAE, image inputs, and working tensors.

Official fileFormatDownload size
FireRed-Image-Edit-1.1-transformer.safetensorsBF16 safetensors40.86 GB
FireRed-Image-Edit-1.1-transformer-bf16.ggufBF16 GGUF40.87 GB
FireRed-Image-Edit-1.1-transformer-q4_k_m.ggufQ4_K_M GGUF13.07 GB
FireRed-Image-Edit-1.1-transformer-q4_1.ggufQ4_1 GGUF12.84 GB
qwen2.5vl-7b-bf16.safetensorsBF16 text encoder16.58 GB
qwen2.5-vl-7b-instruct-q8_0.ggufQ8_0 text encoder8.10 GB
qwen_image_vae.safetensorsVAE0.25 GB

The project repository reports an optimized inference path at about 30 GB VRAM using distillation, quantization, cache, and static compilation together. That is not equivalent to saying every 13 GB GGUF workflow fits in 13 GB VRAM. Record measured peak VRAM on your own GPU and ComfyUI build before treating a configuration as production-safe.

When to choose BF16

Use BF16 when you have enough memory and need a clean reference for output quality, identity consistency, or adapter testing. It is also the safest baseline when diagnosing whether a problem comes from the prompt, workflow, model revision, or quantization.

The transformer is only one part of the workflow. The text encoder, VAE, input images, latent tensors, and attention workspace also consume memory, so a model's on-disk size is not a complete VRAM requirement.

When to choose FP8

FP8 can reduce transformer memory and may run well on hardware and kernels with mature FP8 support. Before using a conversion, verify:

  1. Its base model is exactly FireRedTeam/FireRed-Image-Edit-1.1.
  2. The author documents the quantization method and intended loader.
  3. The file works with the same text encoder, VAE, and workflow revision you plan to use.
  4. Identity, text rendering, and small-detail edits remain acceptable on your own test images.

Do not assume every .safetensors file belongs in the same ComfyUI loader. Follow the conversion author's loader instructions and keep the official workflow unchanged for your first comparison.

When to choose GGUF

GGUF is useful when you need selectable quantization levels and have a compatible ComfyUI-GGUF workflow. Lower-bit variants generally reduce memory further, while higher-bit variants usually preserve more of the BF16 reference behavior. The exact result depends on the conversion and loader, so compare with repeatable inputs rather than relying only on bit labels.

Keep the text encoder and VAE versions aligned with the workflow. Replacing only the transformer is easier to test than changing several components at once.

A practical A/B test

Use the same source image, prompt, seed, resolution, sampler, step count, and adapter state for every format. Compare:

  • face and character identity;
  • small text and decorative typography;
  • edge detail around hair, hands, and products;
  • whether the requested edit occurs without changing protected regions;
  • peak VRAM and total generation time.

Choose the smallest format that passes your real editing cases. A file that merely loads is not automatically the best production choice.