Rift-ai
/

Rift.1-decoder

@@ -1,148 +0,0 @@
----
-license: other
-license_name: rift-non-commercial-license-v1.0
-license_link: ./LICENSE.md
-language:
-- en
-pipeline_tag: image-to-image
-library_name: diffusers
-tags:
-- text-to-image
-- image-editing
-- flux
-- diffusion-single-file
-- vae
-- decoder
-- rift
-- rift-ai
----
-![Comparison Panel](./comparison_panel.jpeg)
-`Rift.1-decoder` is a FLUX-compatible VAE decoder made for the Rift model line. It is designed as a **drop-in decoder component** for compatible FLUX.2-style Diffusers pipelines that use `AutoencoderKLFlux2`. The encoder path remains compatible with FLUX-style latent encoding, while the decoder has been trained as the Rift image reconstruction component.
-The exported Diffusers runtime class remains `AutoencoderKLFlux2` for loader compatibility. The model metadata identifies the architecture as `Rift1Decoder` with model type `rift1_decoder`.
-# **Key Features**
-1. **FLUX-compatible decoder interface** using `AutoencoderKLFlux2`.
-2. **Rift1Decoder metadata** in `config.json` for clear model identity.
-3. **32 latent channels** for compatibility with FLUX.2-style latent spaces.
-4. **512px reconstruction training** with edge and frequency losses for sharper detail retention.
-5. **Single-file artifacts included** for decoder-focused workflows:
-   - `diffusion_pytorch_model.safetensors`
-   - `full_encoder_small_decoder.safetensors`
-   - `small_decoder.safetensors`
-6. Released under the **Rift Non-Commercial License v1.0**.
-Compatible target pipeline family:
-- FLUX.2-style Diffusers pipelines using `AutoencoderKLFlux2`
-- [FLUX.2-klein-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B)
-- [FLUX.2-klein-9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B)
-- [FLUX.2-klein-9b-kv](https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-kv)
-- [FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev)
-# **Comparison**
-| Reference Decoder | Rift1Decoder |
-|:---:|:---:|
-| ![Reference Decoder](./compare_full_decoder.png) | ![Rift1Decoder](./compare_small_decoder.png) |
-# **Detail View**
-![Detail Zoom](./detail_zoom.jpeg)
-# **Usage**
-```shell
-pip install git+https://github.com/huggingface/diffusers.git transformers accelerate torch
-```
-```python
-import torch
-from diffusers import AutoencoderKLFlux2
-vae = AutoencoderKLFlux2.from_pretrained(
-    "Rift-ai/Rift.1-decoder",
-    torch_dtype=torch.bfloat16,
-)
-```
-If using a compatible FLUX.2 pipeline, pass this VAE when loading the pipeline:
-```python
-import torch
-from diffusers import Flux2KleinPipeline, AutoencoderKLFlux2
-device = "cuda"
-dtype = torch.bfloat16
-vae = AutoencoderKLFlux2.from_pretrained("Rift-ai/Rift.1-decoder", torch_dtype=dtype)
-pipe = Flux2KleinPipeline.from_pretrained(
-    "black-forest-labs/FLUX.2-klein-4B",
-    vae=vae,
-    torch_dtype=dtype,
-)
-pipe.enable_model_cpu_offload()
-prompt = "A black cat holding a sign that says 'hello world' in typewriter font"
-image = pipe(
-    prompt=prompt,
-    height=1024,
-    width=1024,
-    guidance_scale=1.0,
-    num_inference_steps=4,
-    generator=torch.Generator(device=device).manual_seed(0),
-).images[0]
-image.save("rift-decoder-output.png")
-```
----
-# **Artifact Files**
-| File | Purpose |
-|:---|:---|
-| `config.json` | Diffusers config with Rift metadata |
-| `diffusion_pytorch_model.safetensors` | Standard Diffusers weights |
-| `full_encoder_small_decoder.safetensors` | Full autoencoder-format weights |
-| `small_decoder.safetensors` | Decoder-only and post-quant-conv weights |
-| `comparison_panel.jpeg` | Full reference/Rift comparison |
-| `compare_full_decoder.png` | Reference decoder reconstruction sample |
-| `compare_small_decoder.png` | Rift decoder reconstruction sample |
-| `detail_zoom.jpeg` | Zoomed detail comparison |
-| `editing.jpg` | Additional visual sample |
----
-# **Limitations**
-- This repository contains a VAE decoder component, not a complete text-to-image model.
-- Visual quality depends on the surrounding diffusion model, scheduler, prompt, latent distribution, and inference settings.
-- The decoder may introduce color shifts, texture smoothing, edge artifacts, or small structural artifacts.
-- Text rendered in generated images may be inaccurate or distorted.
-- Prompt following is handled primarily by the surrounding generation pipeline, not the VAE decoder alone.
-- This model should be evaluated visually and quantitatively before production use.
-# **Out-of-Scope Use**
-This model and its derivatives may not be used outside the scope of the Rift Non-Commercial License v1.0, including for unlawful, fraudulent, defamatory, abusive, exploitative, privacy-invasive, or otherwise harmful purposes.
----
-# **Responsible AI Development**
-Rift.1-decoder should be evaluated as part of a complete image generation or image reconstruction system. A decoder can affect visual fidelity and artifacts, but safety behavior also depends on the text encoder, diffusion transformer, prompt filters, data pipeline, deployment environment, and downstream product policy.
-Users are responsible for applying appropriate safeguards, content review, watermarking or provenance notices where required, and compliance with applicable law.
----
-# **License**
-This model is licensed under the [Rift Non-Commercial License v1.0](./LICENSE.md).
-# **Trademarks & IP**
-This project may contain trademarks or references to third-party projects, products, or services. Use of Rift, Rift-ai, or associated marks in modified versions of this project must not imply sponsorship, endorsement, approval, or official status unless explicitly authorized. Third-party trademarks, intellectual property, and logos remain subject to their respective owners' policies.