<div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/oBQSdNoQtMw" start="0"></iframe></div><h1 id="version-3-(please-read)">VERSION 3 (please read)</h1><h3 id="key-features">🚀 Key Features</h3><ul><li>Multi-Reference Editing: Inject up to 4 reference images to guide your generations.</li><li>Integrated LoRA Stack: Manage up to 4 LoRAs with fingerprint-based no-op detection.</li><li>Advanced Seam Smoothing: Built-in "Patch Model Smoothing" logic to fix bad tiling and textures.</li><li>Expanded Sampler Support: Full flexibility with native ComfyUI samplers and schedulers.</li><li>Seam Visualizer: Heatmap-based analysis to monitor and perfect generation consistency.</li></ul>To UPDATE the nodes (if you already have them):<ol><li>open a command prompt in the "Rebels_HiDream_01_Image_dev_NODES" folder.</li><li>run the command "git pull" to update the nodes.</li></ol><h3 id="then-restart-comfy-and-drag-the-workflow-into-comfy!">then restart comfy and drag the workflow into comfy!</h3>Youtube video tutorial to UPDATE:<a target="_blank" rel="ugc" href="https://youtu.be/iRo-S9oxGe8?si=tSBfVWDgEDqkeVfo">https://youtu.be/iRo-S9oxGe8?si=tSBfVWDgEDqkeVfo</a><h3 id="read-the-entire-description-so-you-dont-miss-a-step."> READ THE ENTIRE DESCRIPTION SO YOU DONT MISS A STEP.</h3>node set to run the HiDream-01 Image Dev GGUF from smthem and the comfy-org bf16 model<h1 id="rebel-hidream-o1-image-dev-(bf16-and-gguf-nodes-for-comfyui)">Rebel HiDream-O1 Image Dev (bf16 and GGUF Nodes for ComfyUI)</h1><h2></h2><h3 id="created-by-rebel-ai">Created by Rebel AI</h3>This repository provides custom ComfyUI nodes to run HiDream-O1-Image-Dev GGUF models locally.HiDream-O1 is a VAE-less, Pixel-Level Unified Transformer. Because it generates raw pixels token-by-token at massive resolutions, running it locally requires careful memory management. These nodes feature upfront dequantization (converting GGUF weights to native PyTorch tensors in system RAM during the load phase) to completely bypass NumPy single-threaded CPU bottlenecks during generation, allowing your GPU to run at maximum efficiency.<h3 id="prerequisites">📦 Prerequisites</h3>Before installing the custom nodes, you need the upstream model code and the weights.Clone the Upstream HiDream-O1 Repo:REQUIRED. NO EXCEPTIONS. WORKFLOW WILL NOT RUN.The nodes rely on the official pipeline logic. Clone this anywhere on your local system:git clone <a target="_blank" rel="ugc" href="https://github.com/HiDream-ai/HiDream-O1-Image.git">https://github.com/HiDream-ai/HiDream-O1-Image.git</a>Download the bf16 Model:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints">https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints</a>Place the bf16 file in CHECKPOINTS: ComfyUI/models/checkpoints/Download the GGUF Model:<a target="_blank" rel="ugc" href="https://huggingface.co/smthem/HiDream-O1-Image-Dev/blob/main/HiDream-O1-Image-Dev-Q6_K.gguf">https://huggingface.co/smthem/HiDream-O1-Image-Dev/blob/main/HiDream-O1-Image-Dev-Q6_K.gguf</a>Place the .gguf file in DIFFUSION_MODELS: ComfyUI/models/diffusion_models/<h3 id="installation">🛠️ Installation</h3>Option 1: ComfyUI Windows PortableNote: Ensure your ComfyUI portable installation is located on a strict local system drive (e.g., C:\ or D:\). Do not install or run these nodes from a OneDrive-synced folder, as it will cause virtual environment and pathing errors.Open a command prompt and navigate to your portable custom nodes directory:cd \ComfyUI_windows_portable\ComfyUI\custom_nodesClone this repository:git clone <a target="_blank" rel="ugc" href="https://github.com/RealRebelAI/Rebels_HiDream-01_Image_Dev_NODES">https://github.com/RealRebelAI/Rebels_HiDream-01_Image_Dev_NODES</a>Install the requirements using the embedded Python environment:cd Rebels_HiDream_01_Image_Dev_NODES"..\..\..\python_embeded\python.exe" -m pip install -r requirements.txtOption 2: Desktop / Standard Python EnvironmentNavigate to your ComfyUI custom nodes directory and lone this repository:cd ComfyUI/custom_nodesgit clone <a target="_blank" rel="ugc" href="https://github.com/RealRebelAI/Rebels_HiDream-01_Image_Dev_NODES">https://github.com/RealRebelAI/Rebels_HiDream-01_Image_Dev_NODES</a>Activate your ComfyUI virtual environment and install the requirements:cd Rebels_HiDream_01_Image_Dev_NODESpip install -r requirements.txt<h3 id="node-documentation">🧩 Node Documentation</h3>Rebel HiDream-O1 Loader (gguf)Loads the GGUF model and performs upfront dequantization.gguf_name: Select your .gguf model from the diffusion_models folder.tokenizer_path: Default is HiDream-ai/HiDream-O1-Image-Dev. It will automatically fetch the tokenizer config from Hugging Face.upstream_repo_path: The absolute local path to where you cloned the HiDream-O1-Image repository in the prerequisites (e.g., C:\Users\name\HiDream-O1-Image).device: Set to cuda for GPU acceleration.offload:aggressive: Heavily utilizes system RAM offloading (Recommended for 8GB VRAM cards like the RTX 3070).balanced: Standard memory splitting.minimal: Keeps most of the model in VRAM.Note: The loader will hang for a moment at 100% while it unpacks the uint8 GGUF bytes into native PyTorch tensors in your system RAM. This is normal and prevents the CPU from bottlenecking your GPU during the actual generation steps.Rebel HiDream-O1 SamplerConnect the model output from the Loader here.steps: 20-30 is the recommended sweet spot.cfg: Keep at 0.0 Higher CFG combined with heavy styling tags can cause "deep-fried" or crushed-shadow artifacts due to the literal pixel-rendering nature of the model.shift: keep at 1.0. Controls the timestep scheduling curve.scheduler_name:default: Standard sampling.flash: Injects specific noise profiles. Note: The noise_scale_start, noise_scale_end, and noise_clip_std parameters only apply if the scheduler is set to flash.⚠️ Known Limitations &amp; BehaviorsResolution Snapping: HiDream-O1 enforces strict token sequence lengths based on its pre-trained position embeddings. If you input a lower resolution (like 512x512 or 1024x1024), the upstream pipeline will automatically "snap" and force the generation to 2048x2048.Compute Times: Because the model renders a raw 2048x2048 image pixel-by-pixel without a VAE, generation times will be significantly longer than standard latent models (like SDXL or Flux).Prompting Style: Avoid stacking heavy texture tags (e.g., "8k resolution, ultra-detailed, gritty textures") unless you want a heavily illustrated look. For photorealism, use clean, simple photographic terms.<h2 id="node-documentation-(bf16-safetensors)">🧩 Node Documentation (BF16 / Safetensors)</h2><h3 id="rebel-hidream-o1-loader-(safetensors)">Rebel HiDream-O1 Loader (Safetensors)</h3>Loads the pure, uncompressed <code>bfloat16</code> safetensors model directly, preserving 100% of the original mathematical precision to eliminate quantization artifacts.<ul><li><code>model_name</code>: Select your <code>.safetensors</code> model from the <code>models/checkpoints/</code> folder.</li><li><code>tokenizer_path</code>: Default is <code>HiDream-ai/HiDream-O1-Image-Dev</code>. It will automatically fetch the tokenizer config from Hugging Face.</li><li><code>upstream_repo_path</code>: The absolute local path to where you cloned the <code>HiDream-O1-Image</code> repository in the prerequisites (e.g., <code>C:\Users\name\HiDream-O1-Image</code>).</li><li><code>device</code>: Set to <code>cuda</code> for GPU acceleration.</li><li><code>offload</code>:<ul><li><code>aggressive</code>: Heavily utilizes system RAM offloading (Recommended for 8GB VRAM cards like the RTX 3070).</li><li><code>balanced</code>: Standard memory splitting.</li><li><code>minimal</code>: Keeps most of the model in VRAM.</li></ul></li></ul>Note: The loader will take a moment to memory-map the massive 16.4GB <code>bfloat16</code> file and balance it across your system RAM and VRAM. Because this workflow uses the native uncompressed weights instead of heavily quantized blocks, the model is much more sensitive to guidance.<h3 id="rebel-hidream-o1-sampler">Rebel HiDream-O1 Sampler</h3>Connect the <code>model</code> output from the Loader here.<ul><li><code>steps</code>: 20-30 is the recommended sweet spot.</li><li><code>cfg</code>: Keep at 0.0 Higher CFG combined with heavy styling tags can cause "deep-fried" or crushed-shadow artifacts due to the literal pixel-rendering nature of the model.</li><li><code>shift</code>: Keep at 1.0. Controls the timestep scheduling curve, delaying fine-detail rendering to accommodate the massive canvas size.</li><li><code>scheduler_name</code>:<ul><li><code>default</code>: Standard sampling.</li><li><code>flash</code>: Injects specific noise profiles. Note: The <code>noise_scale_start</code>, <code>noise_scale_end</code>, and <code>noise_clip_std</code> parameters only apply if the scheduler is set to <code>flash</code>.</li></ul></li></ul><hr /><h2 id="known-limitations-and-behaviors">⚠️ Known Limitations &amp; Behaviors</h2><ul><li>Resolution Snapping: HiDream-O1 enforces strict token sequence lengths based on its pre-trained position embeddings. If you input a lower resolution (like <code>512x512</code> or <code>1024x1024</code>), the upstream pipeline will automatically "snap" and force the generation to 2048x2048.</li><li>Compute Times: Because the model renders a raw 2048x2048 image pixel-by-pixel without a VAE, generation times will be significantly longer than standard latent models (like SDXL or Flux).</li><li>Prompting Style: Avoid stacking heavy texture tags (e.g., "8k resolution, ultra-detailed, gritty textures") unless you want a heavily illustrated look. For photorealism, use clean, simple photographic terms.</li></ul>

Rebels HiDream-01 Image Dev

adds multi-reference editing to the nodes while maintaining T2I capabilities

Files with SHA256: 8289f7971dd86328c92666893757c62a19fc8bc7e541e41a9984541bd8deabd7