FLUX 3 Image Klein — GGUF Quantizations
Community quantizations by Rebel AI of Black Forest Labs' FLUX 3 Image Klein diffusion models, including Base and Step-Distilled variants.
Release status: LIVE. Community GGUF quantizations are available in the Files tab. Download the specific quantization you want to use. Availability varies by file; consult the Files tab for the current releases.
Variants
| Variant | Source checkpoint | Typical use | Starting sampler settings |
|---|---|---|---|
| Base | flux3_image_klein_bf16.safetensors | Quality, control, and experimentation | Euler / Simple, ~50 steps, CFG 3.5 |
| Step-Distilled | flux3_image_klein_sd_bf16.safetensors | Faster generation | Euler / Simple, ~4–16 steps, CFG 1.0 |
Both versions support the FLUX 3 generation/editing workflow family. The settings above are starting points, not GGUF-specific quality guarantees.
GGUF quantization ladder
The release is organized around the following GGUF quantization formats for Base and Step-Distilled. The Files tab lists the formats currently uploaded:
| Format | Intended role |
|---|---|
Q8_0 | Higher-precision GGUF baseline |
Q6_K | High-quality K quant |
Q5_K_M | 5-bit mixed preset |
Q5_K_S | 5-bit smaller preset |
Q5_K | Raw 5-bit K quant |
Q4_K_M | 4-bit mixed preset |
Q4_K_S | 4-bit smaller preset |
Q4_K | Raw 4-bit K quant |
Q4_0 | Standard 4-bit alternative |
Note: Q3_K_S, Q3_K_M, and Q2_K were not supported by the converter used for this ladder. Mixed K presets may contain higher-precision tensors, so the quantization label describes the preset rather than guaranteeing uniform precision across all tensors.
Naming convention
- Base:
flux3_image_klein-<QUANT>.gguf - Step-Distilled:
flux3_image_klein_sd-<QUANT>.gguf
Download links and file sizes are available in the Files tab. Not every format in the ladder is necessarily present at the same time.
ComfyUI setup
Use an up-to-date ComfyUI build with FLUX 3 support and the ComfyUI-GGUF custom node or another GGUF loader supporting FLUX 3. If you encounter missing tensor errors, update the loader and ComfyUI before troubleshooting the model itself.
Put model components in the corresponding directories:
ComfyUI/models/diffusion_models/ # FLUX 3 diffusion GGUFs
ComfyUI/models/text_encoders/ # Qwen3-VL encoder and optional prompt upsampler
ComfyUI/models/vae/ # FLUX 2 VAE
Companion models from the official Comfy checkpoint repository:
qwen3vl_4b_int8_convrot.safetensors— shared text encoderflux2-vae.safetensors— shared VAEqwen3.5_9b_flux3_prompt_upsampler_int8_convrot.safetensors— optional prompt upsampler
The optional 9B upsampler is demanding on low-VRAM systems. Bypass it for initial model testing if necessary.
Official ComfyUI workflows
| Variant | Text-to-image | Image editing |
|---|---|---|
| Base | Workflow JSON | Workflow JSON |
| Step-Distilled | Workflow JSON | Workflow JSON |
The official workflows target the original Comfy safetensors checkpoints. Replace the diffusion model loader with the appropriate GGUF model loader and select your downloaded GGUF. Keep the text encoder and VAE connections unchanged.
Conversion notes and troubleshooting
These community quantizations were produced from the original BF16 Comfy-format diffusion checkpoints, using a GGUF conversion path with FLUX 3 architecture detection. The model tensor names are preserved rather than renamed to an older FLUX architecture layout. Some quantization formats may retain selected tensors at higher precision.
For consistent comparisons, use the same prompt, seed, sampler, resolution, text encoder, VAE, and step/CFG settings when comparing GGUF with BF16. Output quality, memory consumption, and speed can vary by quantization, GPU, and software version. The suggested sampler settings above are starting points, not guarantees.
If you encounter problems, report the GGUF filename, ComfyUI version, GGUF loader version, GPU, sampler settings, and relevant error log. For unexpected image artifacts, include a BF16 comparison where possible.
Credits and licensing
- Original models: Black Forest Labs
- Quantizations / repository: Rebel AI
- ComfyUI: Comfy-Org
- GGUF converter project: molbal/ComfyUI-GGUF
This is an independent community release, not an official Black Forest Labs release. Usage and redistribution remain subject to the upstream model's applicable license and terms. Consult the upstream model card before using or redistributing the weights; this README does not grant additional rights.
Links
Rebel AI on Hugging Face · FLUX 3 Image Klein original · ComfyUI workflows