Open-H3-Regenerate-2K for ComfyUI: refinement nodes, noise controls and comparison examples
Hi,
I put together a small experimental ComfyUI wrapper for NVIDIA’s Open-H3-Regenerate-2K.
IMHO it's a bit like SoL-Refiner when it comes to output, but I decided I'll still set up a node for own tests after initial disappointment.
There are already other ComfyUI options for this like Kijai's LoRA (which I think contains the refiner only?), but this one wraps NVIDIA’s published merged adapters, Sharp curve branches and learned latent upscaler into a few nodes, mainly to make experimentation easier.
Main features:
- Refine existing 24 fps video or native MiniMax-H3 latents.
- Adjustable noise, an upscale-only mode and memory controls.
- Longer clips through independent windows, which can show visible boundaries.
- Example workflows, a test clip and detailed comparison images included.
The attached example is 1344x768 to 2560x1440, using two steps from noise 0.5. It adds texture and definition, though faces, and small details like jewelry, lettering and small patterns can also change. The README in my repo compares upscale-only and noise 0.5 / 0.6 / 0.7.
Tested on Windows with an RTX 5090 (32 GB VRAM) and 128 GB RAM. Smaller cards haven’t been validated; it does consume quite a lot RAM and VRAM, but probably runs at least on an RTX 4090, and you can dial down a few settings to try to make it fit.
I'm Interested to hear how it behaves with other source material.
Repo:
o-l-l-i/ComfyUI-Olm-Open-H3-Regenerate-2K
NVIDIA's official pages for this:
nvlabs.github.io
github.com/NVlabs/Sana/tree/sol-engine/models/Ope…
huggingface.co
24 points · 18 comments
Navigating the Architecture: The Updated Guide to Open-Source Decision Models (October 2026)
The landscape of generative AI has shifted decisively toward open-weight models. Only a few years ago, achieving high-fidelity results typically required relying on a locked-down proprietary API. Today, the situation is entirely different. Current open-weight models are not just competitive; they often outperform closed alternatives in terms of flexibility, prompt adherence, and, most importantly, data privacy. For engineering teams and individual developers, self-hosting is now a viable architectural choice rather than a niche hobbyist project. You gain full ownership of your data, escape the annoyance of arbitrary rate limits, and benefit from predictable, hardware-driven costs.
!Blog Image
The Current Leaderboard Dynamics
To understand where the industry stands, we look toward the Artificial Analysis Text-to-Image Arena. As of October 2, 2026, the board uses the "AA-Image-T2I v2.0" standard, which relies on blind human voting to calculate Elo ratings. The leaderboard has been recalibrated, with FLUX.2 [dev] serving as the anchor point at exactly 1,000.
It is vital to note that some of the most performant models are released under restrictive, non-commercial licenses. While this is great for R&D and internal prototyping, product developers must carefully audit their stack against the following classification of top-tier models.
!Blog Image
Evaluating the Top-Performing Models
1. Qwen-Image-2.1 (Alibaba)
Released on September 20, 2026, Qwen-Image-2.1 jumped immediately to the top of the open-weight rankings with an Elo of 1,036. It is a compact 7B-parameter model utilizing 32 single-stream DiT layers. Its key differentiator is its native ability to generate RGBA images with transparency, eliminating the need for post-processing background removal. It also supports up to 10 reference images, making it a powerhouse for complex multi-image editing workflows.
2. Ideogram 4.0
Ideogram offers a 9.3B-parameter transformer that excels at structured prompts. Rather than relying on fuzzy linguistic interpretation, it allows developers to specify layout, color palettes, and text placement via JSON input. It is an exceptional tool for generating marketing assets, mockups, and typographic designs.
3. FLUX.2 (Black Forest Labs)
FLUX.2 remains the benchmark for image quality. The [dev] variant is a massive 32B-parameter model capable of 4-megapixel output. For those constrained by VRAM, the [klein] 4B version provides a path to high-performance inference on consumer-grade hardware such as the NVIDIA RTX 3090.
4. Ming-Image-0.1-Design (InclusionAI)
Specifically engineered for UI layouts and infographics, Ming-Image-0.1-Design is a specialized tool that punches well above its 6B-parameter weight. It scores a 998 on the Arena and is distributed under the MIT license, making it a rare and powerful choice for commercial applications.
!Blog Image
Essential User Interfaces for Local Deployment
To effectively run these models, you need a robust UI layer. The following are the current industry standards:
- SwarmUI: Ideal for professional environments, this tool supports distributed tasks across multiple GPUs and machines. It is highly optimized for testing prompts across different model variants.
- ComfyUI: The gold standard for power users. Its node-based architecture allows you to create highly complex and reproducible image generation pipelines.
- Forge Neo: A community-maintained fork of the classic WebUI, offering a more accessible, single-page interface for developers who prioritize simplicity and lower memory overhead.
Exposing Local Instances Securely
Running these services locally often leaves you needing to expose them to external collaborators or testing suites. If you are running an interface on your local machine, you do not need to manually configure reverse proxies or expose your firewall. Using Pinggy, you can instantly tunnel your local port to a public HTTPS URL with one simple command.
# Example for exposing ComfyUI at port 8188
ssh -p 443 -R0:localhost:8188 free.pinggy.io
This method is perfect for rapid prototyping and sharing results with clients without the complexity of traditional cloud hosting configurations.
Deep-Dive Troubleshooting and Architecture
When deploying these models, memory management is the primary bottleneck. For models like Cosmos3, which require data-center hardware (8xH100), you must ensure that your environment supports high-bandwidth memory (HBM) and proper thermal management. In contrast, models like Sana are specifically optimized for efficiency, allowing you to run them on consumer laptops.
If you encounter out-of-memory (OOM) errors, consider the following strategies:
1. Quantization: Use fp8 or nf4 checkpoints to reduce VRAM footprint.
2. Model Offloading: Utilize the device_map="auto" argument in the Hugging Face diffusers library to spill excess tensors to system RAM.
3. Tile-based Generation: For high-resolution outputs, use tiling to process image segments sequentially, effectively bypassing native VRAM limits.
Future Outlook
As we move into 2027, the gap between commercial APIs and open-weight models will likely continue to shrink. The emergence of "agentic" image generation, where models autonomously verify and refine their own outputs, is the next major hurdle. We recommend starting with a flexible setup using ComfyUI and experimenting with models like Qwen-Image-2.1 for research or Ming-Image-0.1-Design for commercial production.
Reference
- Best Free & Open-Source AI Image Generators to Self-Host
- Artificial Analysis Arena
- Hugging Face
5 reactions
Mentions
Open-H3-Regenerate-2K for ComfyUI: refinement nodes, noise controls and comparison examples Hi, I put together a small experimental ComfyUI wrapper for NVIDIA’s Open-H3-Regenerate-2K. IMHO it's a bit like SoL-Refiner when it comes to output, but I decided I'll still set up a node for own tests after initial disappointment. There are already other ComfyUI options for this like Kijai's LoRA (which I think contains the refiner only?), but this one wraps NVIDIA’s published merged adapters, Sharp curve branches and learned latent upscaler into a few nodes, mainly to make experimentation easier. Main features: - Refine existing 24 fps video or native MiniMax-H3 latents. - Adjustable noise, an upscale-only mode and memory controls. - Longer clips through independent windows, which can show visible boundaries. - Example workflows, a test clip and detailed comparison images included. The attached example is 1344x768 to 2560x1440, using two steps from noise 0.5. It adds texture and definition, though faces, and small details like jewelry, lettering and small patterns can also change. The README in my repo compares upscale-only and noise 0.5 / 0.6 / 0.7. Tested on Windows with an RTX 5090 (32 GB VRAM) and 128 GB RAM. Smaller cards haven’t been validated; it does consume quite a lot RAM and VRAM, but probably runs at least on an RTX 4090, and you can dial down a few settings to try to make it fit. I'm Interested to hear how it behaves with other source material. Repo: o-l-l-i/ComfyUI-Olm-Open-H3-Regenerate-2K NVIDIA's official pages for this: nvlabs.github.io github.com/NVlabs/Sana/tree/sol-engine/models/Ope… huggingface.co
Regenerate Minimax H3 videos in 2K in just 3 steps. It does tend to change details and faces though. github.com/NVlabs/Sana/tree/sol-engine/models/Ope… t.co
🚀 Open-H3-Regenerate-2K: 3-Step Video Regeneration, Native Output up to 4K 📝 Blog: nvlabs.github.io 💻 Code: github.com/NVlabs/Sana/tree/sol-engine/models/Ope… 🤗 HF Weights: huggingface.co t.co
5秒视频，1.65秒就生成完了。 英伟达开源 Sol-H3，专门给 MiniMax H3 视频模型提速。 这是在8张 B300、1344×768分辨率下跑出来的成绩，生成的视频还带同步音频。 🔗 github.com/NVlabs/Sana/tree/sol-engine t.co
英伟达开源Sol-H3： MiniMax H3提速11倍， 5秒视频1.65秒完成 github.com/NVlabs/Sana/tree/sol-engine
Navigating the Architecture: The Updated Guide to Open-Source Decision Models (October 2026) The landscape of generative AI has shifted decisively toward open-weight models. Only a few years ago, achieving high-fidelity results typically required relying on a locked-down proprietary API. Today, the situation is entirely different. Current open-weight models are not just competitive; they often outperform closed alternatives in terms of flexibility, prompt adherence, and, most importantly, data privacy. For engineering teams and individual developers, self-hosting is now a viable architectural choice rather than a niche hobbyist project. You gain full ownership of your data, escape the annoyance of arbitrary rate limits, and benefit from predictable, hardware-driven costs. !Blog Image The Current Leaderboard Dynamics To understand where the industry stands, we look toward the Artificial Analysis Text-to-Image Arena. As of October 2, 2026, the board uses the "AA-Image-T2I v2.0" standard, which relies on blind human voting to calculate Elo ratings. The leaderboard has been recalibrated, with FLUX.2 [dev] serving as the anchor point at exactly 1,000. It is vital to note that some of the most performant models are released under restrictive, non-commercial licenses. While this is great for R&D and internal prototyping, product developers must carefully audit their stack against the following classification of top-tier models. !Blog Image Evaluating the Top-Performing Models 1. Qwen-Image-2.1 (Alibaba) Released on September 20, 2026, Qwen-Image-2.1 jumped immediately to the top of the open-weight rankings with an Elo of 1,036. It is a compact 7B-parameter model utilizing 32 single-stream DiT layers. Its key differentiator is its native ability to generate RGBA images with transparency, eliminating the need for post-processing background removal. It also supports up to 10 reference images, making it a powerhouse for complex multi-image editing workflows. 2. Ideogram 4.0 Ideogram offers a 9.3B-parameter transformer that excels at structured prompts. Rather than relying on fuzzy linguistic interpretation, it allows developers to specify layout, color palettes, and text placement via JSON input. It is an exceptional tool for generating marketing assets, mockups, and typographic designs. 3. FLUX.2 (Black Forest Labs) FLUX.2 remains the benchmark for image quality. The [dev] variant is a massive 32B-parameter model capable of 4-megapixel output. For those constrained by VRAM, the [klein] 4B version provides a path to high-performance inference on consumer-grade hardware such as the NVIDIA RTX 3090. 4. Ming-Image-0.1-Design (InclusionAI) Specifically engineered for UI layouts and infographics, Ming-Image-0.1-Design is a specialized tool that punches well above its 6B-parameter weight. It scores a 998 on the Arena and is distributed under the MIT license, making it a rare and powerful choice for commercial applications. !Blog Image Essential User Interfaces for Local Deployment To effectively run these models, you need a robust UI layer. The following are the current industry standards: - SwarmUI: Ideal for professional environments, this tool supports distributed tasks across multiple GPUs and machines. It is highly optimized for testing prompts across different model variants. - ComfyUI: The gold standard for power users. Its node-based architecture allows you to create highly complex and reproducible image generation pipelines. - Forge Neo: A community-maintained fork of the classic WebUI, offering a more accessible, single-page interface for developers who prioritize simplicity and lower memory overhead. Exposing Local Instances Securely Running these services locally often leaves you needing to expose them to external collaborators or testing suites. If you are running an interface on your local machine, you do not need to manually configure reverse proxies or expose your firewall. Using Pinggy, you can instantly tunnel your local port to a public HTTPS URL with one simple command. # Example for exposing ComfyUI at port 8188 ssh -p 443 -R0:localhost:8188 free.pinggy.io This method is perfect for rapid prototyping and sharing results with clients without the complexity of traditional cloud hosting configurations. Deep-Dive Troubleshooting and Architecture When deploying these models, memory management is the primary bottleneck. For models like Cosmos3, which require data-center hardware (8xH100), you must ensure that your environment supports high-bandwidth memory (HBM) and proper thermal management. In contrast, models like Sana are specifically optimized for efficiency, allowing you to run them on consumer laptops. If you encounter out-of-memory (OOM) errors, consider the following strategies: 1. Quantization: Use fp8 or nf4 checkpoints to reduce VRAM footprint. 2. Model Offloading: Utilize the device_map="auto" argument in the Hugging Face diffusers library to spill excess tensors to system RAM. 3. Tile-based Generation: For high-resolution outputs, use tiling to process image segments sequentially, effectively bypassing native VRAM limits. Future Outlook As we move into 2027, the gap between commercial APIs and open-weight models will likely continue to shrink. The emergence of "agentic" image generation, where models autonomously verify and refine their own outputs, is the next major hurdle. We recommend starting with a flexible setup using ComfyUI and experimenting with models like Qwen-Image-2.1 for research or Ming-Image-0.1-Design for commercial production. Reference - Best Free & Open-Source AI Image Generators to Self-Host - Artificial Analysis Arena - Hugging Face
StreamRender-H3: Code-driven Streaming Render Runtime Demo: nvlabs.github.io Code: github.com/NVlabs/Sana/tree/sol-engine/models/str… via：@xieenze_jr