realrebelai/Fibo-1.5_GGUFs

Fibo 1.5 --- GGUF & INT8 Quantizations

Quantized versions of BRIA AI Fibo 1.5 for local inference and ComfyUI.

This repository provides a full GGUF quantization ladder plus a native ComfyUI INT8 version of the Fibo 1.5 transformer. The goal is to make the 8B Fibo 1.5 image-generation model more practical on consumer hardware while retaining as much of the original model quality as possible.

ComfyUI Nodes

These models are supported by the custom ComfyUI Fibo 1.5 nodes:

https://github.com/RealRebelAI/ComfyUI_Fibo-1.5/tree/main

Install the repository into:

ComfyUI/custom_nodes/ComfyUI_Fibo-1.5/

Then restart ComfyUI.

The custom nodes provide:

• —Fibo GGUF transformer loading
• —Native Fibo INT8 diffusion-model loading
• —SmolLM3 text-encoder loading
• —Fibo-specific text conditioning
• —Per-block Fibo caption conditioning
• —Native 48-channel Fibo latents
• —Wan/Fibo VAE loading and decoding
• —Portable tokenizer and configuration files
• —Fibo's few-step distilled generation workflow

The required Fibo configuration and tokenizer files are bundled directly inside the custom-node repository. They do not need to be downloaded separately.

Original Model

These quantizations are derived from:

BRIA AI --- Fibo 1.5

https://huggingface.co/briaai/Fibo-1.5

The Hugging Face model-card metadata in this repository explicitly declares:

base_model: briaai/Fibo-1.5
base_model_relation: quantized

so the Hub can associate this repository with the original Fibo 1.5 model as a quantized derivative.

Fibo 1.5 uses an approximately 8B-parameter diffusion transformer with a SmolLM3-based text encoder, Fibo-specific multi-layer conditioning, a Wan-based VAE, structured prompting, and a distilled few-step generation workflow.

Available Quantizations

GGUF

The intended GGUF ladder is:

Quantization Description -------------- ------------------------------------ Q8_0 Highest-fidelity GGUF quantization Q6_K High-quality quantized option Q5_K_M Strong quality/size balance Q4_K_M Balanced Q4 option Q4_K_S Smaller/more aggressive Q4 Q3_K_M Aggressive compression Q2_K Maximum compression / experimental

Each GGUF tier is produced directly from the high-precision transformer source rather than requantizing another lower-precision GGUF.

High-precision source
 ├── Q8_0
 ├── Q6_K
 ├── Q5_K_M
 ├── Q4_K_M
 ├── Q4_K_S
 ├── Q3_K_M
 └── Q2_K

Native INT8

A native ComfyUI INT8 transformer is also provided:

Fibo-1.5-INT8.safetensors

This uses native ComfyUI int8_tensorwise quantization rather than GGUF.

The transformer body is stored in INT8 while selected Fibo-sensitive components remain at source precision.

Protected components:

caption_projection
context_embedder
x_embedder
time_embed
norm_out
proj_out

The conversion produced:

BF16 source: 16.57 GB
INT8:        10.28 GB

Sampled INT8 transformer matrices measured:

Mean relative L2 error:  0.853%
Worst sampled error:     0.857%

Example sampled layers:

0.852%  single_transformer_blocks.0.attn.to_k.weight
0.851%  single_transformer_blocks.0.attn.to_q.weight
0.856%  single_transformer_blocks.0.attn.to_v.weight
0.852%  single_transformer_blocks.0.norm.linear.weight
0.851%  single_transformer_blocks.0.proj_mlp.weight
0.856%  single_transformer_blocks.1.attn.to_k.weight
0.850%  single_transformer_blocks.1.attn.to_q.weight
0.857%  single_transformer_blocks.1.attn.to_v.weight

These are conversion-domain measurements and are not a direct prediction of image quality. Controlled image-generation comparisons are still the appropriate way to evaluate practical fidelity.

Requirements

The quantized files in this repository contain the Fibo transformer. A complete workflow additionally requires:

• —ComfyUI
• —ComfyUI Fibo 1.5 custom nodes
• —Compatible Fibo/SmolLM3 text encoder
• —Compatible Fibo/Wan VAE
• —City96 ComfyUI-GGUF for GGUF models

Custom Node Configuration

The Fibo custom-node repository includes its required configuration and tokenizer files directly inside the node:

ComfyUI/
└── custom_nodes/
    └── ComfyUI_Fibo-1.5/
        ├── nodes.py
        ├── fibo_model.py
        ├── fibo_runtime.py
        └── config/
            ├── vae/
            │   └── config.json
            ├── text_encoder/
            │   └── config.json
            └── tokenizer/
                └── ...

You do not need to manually place these configuration files elsewhere in ComfyUI.

The nodes resolve them relative to their own installation directory. Do not remove or relocate the included config/ folder.

Text Encoder

Fibo 1.5 uses a SmolLM3-based text encoder.

Place the compatible merged Fibo text encoder under:

ComfyUI/models/text_encoders/

Example:

ComfyUI/
└── models/
    └── text_encoders/
        └── Fibo-1.5-text-encoder-BF16.safetensors

Use:

Fibo Text Encoder Loader

followed by:

Fibo Text Encode

VAE

Fibo uses a Wan-based VAE with a native 48-channel latent representation.

Place the compatible VAE weights under:

ComfyUI/models/vae/

Use:

Fibo VAE Loader

Do not substitute a standard Stable Diffusion or Flux latent. The custom workflow provides:

Fibo Empty Latent (48ch)

with the correct Fibo latent layout.

Installing the ComfyUI Nodes

Install:

https://github.com/RealRebelAI/ComfyUI_Fibo-1.5/tree/main

into:

ComfyUI/custom_nodes/ComfyUI_Fibo-1.5/

For GGUF inference, also install City96 ComfyUI-GGUF:

https://github.com/city96/ComfyUI-GGUF

Restart ComfyUI after installation.

Model Placement

GGUF

Place .gguf files in:

ComfyUI/models/diffusion_models/

or:

ComfyUI/models/unet/

Then use:

Fibo GGUF Loader

The custom Fibo runtime constructs the Fibo transformer architecture while City96's GGUF/GGML operations provide quantized tensor execution.

INT8

Place:

Fibo-1.5-INT8.safetensors

in:

ComfyUI/models/diffusion_models/

or:

ComfyUI/models/unet/

Then use:

Fibo INT8 Diffusion Model Loader

Do not load the INT8 .safetensors through the GGUF loader. The INT8 checkpoint uses native ComfyUI quantized tensors and has its own diffusion-model loader.

Included Fibo Nodes

The custom-node repository provides these primary nodes:

Fibo GGUF Loader
Fibo INT8 Diffusion Model Loader
Fibo Text Encoder Loader
Fibo Text Encode
Fibo Empty Latent (48ch)
Fibo VAE Loader

Basic GGUF Workflow

Fibo GGUF Loader ────────────────┐
                                 │
Fibo Text Encoder Loader         │
          ↓                      │
Fibo Text Encode ────────────────┤
                                 ↓
Fibo Empty Latent (48ch) → KSampler
                                 ↓
                         Fibo VAE Loader
                                 ↓
                            Save Image

Basic INT8 Workflow

For INT8, replace:

Fibo GGUF Loader

with:

Fibo INT8 Diffusion Model Loader

The remainder of the workflow stays the same:

Fibo INT8 Diffusion Model Loader ─┐
                                  │
Fibo Text Encoder Loader          │
          ↓                       │
Fibo Text Encode ─────────────────┤
                                  ↓
Fibo Empty Latent (48ch) → KSampler
                                  ↓
                          Fibo VAE Loader
                                  ↓
                             Save Image

Recommended Generation Settings

Fibo 1.5 is a distilled few-step model.

A good starting point:

Steps: 4–6
CFG: 1.0
Sampler: Euler

CFG

Use:

CFG = 1.0

Fibo 1.5 is intended to operate without conventional classifier-free-guidance amplification. High CFG values commonly used with older Stable Diffusion models are not the intended Fibo workflow.

Prompting

Prompt structure matters significantly with Fibo.

Fibo can generate from ordinary natural-language prompts, but detailed structured JSON-style prompts can provide substantially more explicit scene information.

Useful structured fields include:

• —Scene description
• —Subjects
• —Object placement
• —Anatomy
• —Materials
• —Environment
• —Lighting
• —Camera
• —Composition
• —Style
• —Fine details

Structured Prompt Example

{
  "short_description": "A realistic tabby cat sitting beside a window in warm afternoon sunlight.",
  "objects": [
    {
      "object_id": "cat",
      "type": "animal",
      "description": "A realistic brown tabby cat with detailed striped fur, amber eyes, visible whiskers, natural paws and anatomically correct feline proportions.",
      "position": "Centered on a wooden windowsill."
    }
  ],
  "lighting": {
    "conditions": "Warm natural late-afternoon sunlight entering through the window."
  },
  "style_medium": "Photography",
  "artistic_style": "Photorealistic naturalism."
}

Comparing Quantizations

Keep all generation variables identical when comparing quantization tiers:

Prompt
Seed
Resolution
Steps
CFG
Sampler
Scheduler

A useful comparison ladder is:

BF16
  │
INT8
  │
Q8_0
  │
Q6_K
  │
Q5_K_M
  │
Q4_K_M
  │
Q4_K_S
  │
Q3_K_M
  │
Q2_K

Changing the seed can produce significant differences in composition, anatomy, object placement, typography, fine detail, and lighting that are unrelated to quantization.

Quantization and Image Quality

Quantization may affect difficult generation features such as:

• —Fine textures
• —Skin and fur detail
• —Complex anatomy
• —Hands and fingers
• —Small objects
• —Typography and exact text
• —Structured prompt adherence
• —Subtle lighting
• —Material appearance
• —Very small scene details

These are also inherently difficult areas for generative image models at high precision. A single imperfect generation should therefore not automatically be attributed to quantization.

Controlled A/B testing is strongly recommended.

GGUF vs INT8

GGUF and native INT8 use different runtime strategies.

A larger INT8 checkpoint is not automatically faster than a smaller GGUF model.

Performance depends on:

• —GPU architecture
• —Available VRAM
• —System RAM
• —Quantization format
• —Runtime kernels
• —Model offloading
• —Resolution
• —Activation memory
• —Compute dtype

On low-VRAM GPUs, a smaller Q4/Q5 GGUF may run faster than the larger INT8 checkpoint because more of the transformer can remain resident in available memory.

Choosing a GGUF Quant

Q8_0

Use when maximum GGUF fidelity is the priority and memory/storage are less important.

Q6_K

A high-quality quantized option for users who want to stay relatively close to high precision.

Q5KM

A strong middle ground between model size and fidelity.

Q4KM

A practical starting point when reducing memory/storage while trying to retain useful image quality.

Q4KS

A more aggressive Q4 tier when additional size reduction is required.

Q3KM

Aggressive compression for lower-memory systems or experimentation.

Q2_K

Maximum-compression experimental tier. Compare carefully against higher quantizations.

Low-VRAM Systems

Fibo 1.5 remains a large model even after quantization.

On low-VRAM GPUs, ComfyUI may offload portions of the transformer between GPU and system memory. This can substantially affect generation speed.

A smaller quant can therefore outperform a theoretically faster numeric format simply because it reduces memory pressure and model movement.

48-Channel Latent

Fibo requires its native latent representation.

The custom nodes provide:

Fibo Empty Latent (48ch)

with:

Channels: 48
Spatial downscale: 16x

Using an incompatible latent format will result in incorrect tensor shapes or failed generation.

Text Conditioning

The custom Fibo text-encoder implementation follows Fibo's multi-layer conditioning design.

SmolLM3 hidden states are used to create:

• —Main text embeddings
• —Per-transformer-block text conditioning
• —Attention masks

The final two SmolLM3 hidden states are concatenated for the main Fibo context representation. Additional hidden states are supplied to Fibo's per-block caption-projection layers.

This differs from diffusion architectures that consume only a single final text-encoder hidden state.

Troubleshooting

Model does not appear

Make sure the transformer is located in:

ComfyUI/models/diffusion_models/

or:

ComfyUI/models/unet/

Then restart ComfyUI or refresh its model list.

GGUF will not load

Check that:

1. 1.City96 ComfyUI-GGUF is installed.
2. 2.The Fibo custom nodes are installed.
3. 3.You are using Fibo GGUF Loader.
4. 4.The GGUF was produced for the Fibo architecture.

Do not treat Fibo as a Flux checkpoint simply because some tooling may share concepts or formats.

INT8 will not load

Check that:

1. 1.The model is the native Fibo INT8 .safetensors checkpoint.
2. 2.You are using Fibo INT8 Diffusion Model Loader.
3. 3.Your Fibo custom nodes are up to date.

Do not use Fibo GGUF Loader for the INT8 .safetensors.

Tokenizer not found

Verify:

ComfyUI_Fibo-1.5/config/tokenizer/

exists inside the custom node.

Text encoder config not found

Verify:

ComfyUI_Fibo-1.5/config/text_encoder/config.json

exists.

VAE config not found

Verify:

ComfyUI_Fibo-1.5/config/vae/config.json

exists.

VAE decoding is slow

Wan-based VAE decoding can be memory intensive. On lower-VRAM hardware, model unloading and GPU memory pressure can make VAE decoding substantially slower.

This does not by itself indicate that the transformer quantization is broken.

Current Scope

The current project focuses on Fibo 1.5 inference in ComfyUI using:

• —GGUF transformers
• —Native INT8 transformers
• —Fibo text conditioning
• —Fibo 48-channel latents
• —Wan/Fibo VAE decoding

Additional quantization formats can be explored separately.

Credits

BRIA AI

Fibo and Fibo 1.5 were developed by BRIA AI.

Original model:

https://huggingface.co/briaai/Fibo-1.5

All credit for the original Fibo architecture, model weights, and training belongs to BRIA AI and its contributors.

ComfyUI Fibo 1.5 Nodes

Custom ComfyUI integration:

https://github.com/RealRebelAI/ComfyUI_Fibo-1.5/tree/main

City96 --- ComfyUI-GGUF

GGUF loading uses the ComfyUI-GGUF ecosystem developed by City96:

https://github.com/city96/ComfyUI-GGUF

Hugging Face Diffusers

The upstream Fibo pipeline and architecture are available through the Hugging Face Diffusers ecosystem.

License

These files are derivative quantizations of BRIA AI Fibo 1.5.

The original model and its weights remain subject to BRIA AI's applicable Fibo license and usage terms.

Quantization does not replace, remove, or modify the license of the original model.

Users are responsible for reviewing and complying with the current upstream license before using, redistributing, publishing derivative weights, commercially deploying, or otherwise distributing these files.

Refer to the official model repository for the authoritative current licensing terms:

https://huggingface.co/briaai/Fibo-1.5

This quantization repository and the custom ComfyUI implementation are unofficial community projects and are not affiliated with or endorsed by BRIA AI.

Links

Original BRIA Fibo 1.5

https://huggingface.co/briaai/Fibo-1.5

ComfyUI Fibo 1.5 Custom Nodes

https://github.com/RealRebelAI/ComfyUI_Fibo-1.5/tree/main

City96 ComfyUI-GGUF

https://github.com/city96/ComfyUI-GGUF