ComfyUI-HunyuanImage3 Introduction
ComfyUI-HunyuanImage3 brings Tencent’s HunyuanImage-3.0 image-generation models directly into ComfyUI. You can create images from text, edit existing artwork, combine reference images, and optionally let the model expand your prompt before generating.
The extension works with familiar ComfyUI tools, including KSampler, VAE Decode, and ComfyUI’s own memory management. It is a native integration, rather than a wrapper around a separate generation pipeline.
Although HunyuanImage-3.0 is an 80-billion-parameter model, the extension makes it practical to run on a single consumer NVIDIA GPU by keeping most model data in system memory and transferring it to the GPU when needed.
The recommended starting point is Instruct-Distil with W4A8 weights: it supports text-to-image, editing, multi-image fusion, and prompt rewriting, with just eight sampling steps.
Instruct-Distil, Instruct, and Base compared across five prompts
The author’s comparison uses W4A8 weights at 1024 × 1024 on one RTX 4090. Each prompt uses the same seed across models. The first image in each row also includes model-loading time.
What you can do
- Develop artwork from written ideas: describe a subject, setting, style, or composition.
- Transform existing images: for example, “Turn this photo into a watercolor.”
- Combine up to three references: take a subject from one image and place it in the setting of another.
- Expand short prompts: ask the model to turn a brief idea into a more detailed visual description.
- Choose between speed and flexibility: use the eight-step distilled model or the full Instruct and Base models.
- Accelerate longer renders: enable Spectrum when you can accept some changes from the unaccelerated result. Explore the author’s side-by-side comparison gallery →
The gallery compares all three models, all three weight formats, and Spectrum enabled versus disabled.
How ComfyUI-HunyuanImage3 Works
A large model with a team of specialists
HunyuanImage-3.0 uses a mixture-of-experts design. Think of it as a large studio containing many specialists: the entire studio is available, but only selected specialists handle each piece of work.
The model has 64 experts and 80 billion parameters in total, with roughly 13 billion active for a token. This provides a large overall capacity without using every parameter for every calculation.
System memory acts as the storage room
The whole model does not need to fit inside your graphics card.
Instead:
- The model’s weights—the stored data it uses to generate images—are held in system RAM.
- ComfyUI transfers the needed portions into GPU memory, also called VRAM.
- While one layer is working, the extension can prepare the next layer’s weights.
- This repeats throughout generation. This is why a 12 GB or 16 GB GPU can run the recommended model, but a small GPU does not remove the need for substantial system RAM. Transfer speed between RAM and the GPU also matters.
From your prompt to a finished image
A typical text-to-image workflow follows this path:
Load model → encode prompt → sample with KSampler → decode with the VAE → save image
- The encoder prepares your prompt as instructions the model can use.
- KSampler progressively forms the image over a number of steps.
- The VAE converts the model’s compact internal image representation, called a latent, into a visible image. For image editing, Image Encode also prepares your reference images. For prompt rewriting, the model first generates an expanded description and then uses that description to draw.
Why smaller weight formats help
Quantization stores model data more compactly. You can think of it as using a lighter edition of the same reference library.
The recommended W4A8 format reduces storage and transfer demands substantially. It can change image details or composition compared with the original weights, but a numerical difference does not automatically mean lower artistic quality.
ComfyUI-HunyuanImage3 Requirements and Model Files
Practical hardware requirements
| Resource | What to expect |
|---|---|
| GPU | An NVIDIA GPU with at least 12 GB VRAM. The author tested RTX 4090 and RTX 3090 hardware, including constrained 16 GB and 12 GB VRAM tests. |
| System RAM | Substantial free RAM is essential. With W4A8 loaded, ComfyUI used about 50 GB in the author’s test on a 188 GB system. Leave additional room for your operating system and other applications. |
| Storage | Approximately 45 GB for the recommended W4A8 checkpoint, plus supporting files. A fast SSD helps initial loading. |
| ComfyUI | A current version with the required native quantized expert-bank support. The author specifies late September 2026 or newer and tested dynamic VRAM streaming through comfy-aimdo. |
Tencent’s reference pipeline lists much larger multi-GPU requirements. Those figures describe the upstream implementation, not this extension’s quantized, RAM-streaming setup.
Files used by your workflow
Each supplied workflow includes a Read me note with the required files and download links.
| File | ComfyUI folder | Purpose |
|---|---|---|
| hunyuan_image_3_<model>_w4a8.safetensors | models/diffusion_models/ | Recommended main model checkpoint |
| hunyuan_image_3_vae_fp16.safetensors | models/vae/ | Shared image decoder for all three models |
| hunyuan_image_3_<model>_siglip2_so400m_naflex.safetensors | models/clip_vision/ | Vision encoder for image editing and reference-image workflows |
| hunyuan_image_3_<model>_cot_head.safetensors | models/diffusion_models/ | Additional text-generation component for prompt rewriting; approximately 1 GB |
Use the model-specific vision and rewriting files listed in your chosen workflow. The extension already includes the model configuration and tokenizer files.
For ordinary text-to-image without rewriting, you do not need the vision encoder or cot_head file.
ComfyUI-HunyuanImage3 Models
Choosing a model
| Model | Supported tasks | Recommended starting settings | When to choose it |
|---|---|---|---|
| Instruct-Distil | Text-to-image, editing, fusion, prompt rewriting | 8 steps; KSampler cfg: 1.0; Guidance node 2.5 | Everyday artwork, quick iterations, and editing |
| Instruct | Text-to-image, editing, fusion, prompt rewriting | 50 steps; cfg: 2.5 | The full instruction-tuned model, with classic CFG and an adjustable step count |
| Base | Text-to-image | 50 steps; cfg: 5.0 | Exploring the pretrained model’s output, including strong photographic detail |
All three use the same GPU and memory requirements for a given weight format. Instruct-Distil is faster because it needs fewer steps—not because its checkpoint is much smaller.
Instruct-Distil
ComfyUI-ready weights · Tencent model card
This is the default and recommended first choice. It retains the Instruct model’s editing and rewriting capabilities while generating with eight steps.
Use it when you want to explore multiple prompts or seeds without waiting several minutes for each image.
Instruct
ComfyUI-ready weights · Tencent model card
This is the full instruction-tuned model. It supports the same main creative tasks as Instruct-Distil, but the supplied workflow uses 50 steps and classic classifier-free guidance, or CFG.
Choose it when you specifically want to explore the full model’s rendering behavior or sampling-step flexibility. More steps should not be interpreted as a guarantee that every result will look better.
Base
ComfyUI-ready weights · Tencent model card
Base is the pretrained text-to-image model. It is useful for exploring its photographic detail and visual character.
Base does not support this extension’s image-editing or built-in prompt-rewriting workflows. Write the visual details you need directly into your prompt.
Choosing a weight format
Comparison of bf16, int8, and W4A8 at the same prompt and seed
Each row keeps the prompt and seed fixed while changing the weight format. The strips beneath the images show fine-detail crops.
| Format | Approximate checkpoint size | Performance | Best use | |---|---:|---|---| | W4A8 | 44 GiB | Fastest | Recommended for everyday use; 4-bit weights and 8-bit activations | | int8 ConvRot | 76 GiB | About 1.9× slower per step than W4A8 in the reported tests | Comparing a less-compressed version that is closer numerically to the original | | bf16 | 150 GiB | Slowest and highly dependent on available RAM | Unquantized reference comparisons |
The bf16 checkpoint is bit-identical to Tencent’s original weights. In the quantized versions, components that are not quantized—such as embeddings, normalization components, and the router—remain byte-identical to bf16. The author measured image similarity to bf16 at 0.844 SSIM for int8 and 0.738 for W4A8. SSIM measures how closely two images match, not which image is more attractive. Even a small weight change can move an object or alter a composition at the same seed.
ComfyUI-HunyuanImage3 First Image Workflow
1. Open a supplied workflow
Drag a workflow from workflows/ into ComfyUI.
| Workflow file | Purpose |
|---|---|
| hunyuan_image_3_instruct_distil_txt2img.json | Recommended first text-to-image workflow |
| hunyuan_image_3_instruct_distil_img2img.json | Instruct-Distil editing with one to three images |
| hunyuan_image_3_instruct_txt2img.json | Full Instruct text-to-image |
| hunyuan_image_3_instruct_img2img.json | Full Instruct editing with one to three images |
| hunyuan_image_3_base_txt2img.json | Base text-to-image |
2. Select your checkpoint
In Load HunyuanImage 3.0 Model, select the main model file.
The loader identifies Base, Instruct, or Instruct-Distil from the weights. You do not need to set a separate model type.
3. Write your prompt
Describe the subject and the visual decisions that matter to you. For example:
A ceramic fox figurine on a wooden desk, soft morning window light, warm neutral colors, shallow depth of field, product photography.
Both encoders use the selected model’s built-in system prompt. An optional custom_system_prompt input lets you replace it, but the supplied behavior is the simpler starting point.
4. Keep the supplied sampling settings
Always begin with:
- Sampler:
euler - Scheduler:
simple - Steps and CFG: the values for your model in the model table above
- Instruct-Distil Guidance:
2.5The encoders automatically produce the model’s own empty negative conditioning. There is no separate negative-prompt text box to fill in these workflows.
5. Choose an appropriate size and queue
Use HunyuanImage 3.0 Resolutions to select one of the model’s 37 native image shapes.
The model was trained at approximately one megapixel, with shapes ranging from 2048 × 512 to 512 × 2048. Native sizes are the most straightforward place to begin.
The first image includes loading work, so judge ongoing performance from later images with the same model already in memory.
ComfyUI-HunyuanImage3 Features
Text-to-image generation
Use HunyuanImage 3.0 Text Encode to turn your written description into the positive and negative conditioning needed by KSampler.
You can explore different compositions by changing the image seed. A seed is the starting random pattern for generation; keeping it fixed makes setting comparisons easier.
Image editing
Use an _img2img workflow or HunyuanImage 3.0 Image Encode with:
- An Instruct or Instruct-Distil checkpoint.
- Your input image.
- The shared VAE.
- The matching vision-encoder file.
- A written editing instruction. Example:
Turn this portrait into a watercolor illustration while keeping the person’s pose.
Examples of portrait editing and editing with prompt rewriting
The author’s examples use Instruct-Distil W4A8 at eight steps. The comparison gallery includes the instructions used.
Two controls adjust how strongly reference images influence generation:
| Control | Default | Role |
|---|---:|---|
| vit_strength | 1.0 | Adjusts the influence of the vision-encoder representation |
| latent_strength | 1.0 | Adjusts the influence of the latent-image representation |
Treat these as reference-influence controls, not guaranteed preservation sliders. Start at the defaults and change one at a time so you can see which part of the result changes.
Multi-image fusion
Connect up to three images to Image Encode. A new image socket appears when you connect an input.
Refer to each image by order:
Put the cat from the first image on the sofa in the second image, using the color palette of the third image.
Each reference is resized to approximately one megapixel. Adding references makes every sampling step slower because the model reads them throughout generation.
The output dimensions come from the encoder’s width and height inputs. Supplied editing workflows follow the first image’s shape at approximately one megapixel; use the Resolutions node when you want a fixed output shape.
Reference-processing compatibility
reference_vit_padding is enabled by default. It pads the vision-encoder processing to the checkpoint’s 1024 tower slots, matching Tencent’s reference processor.
- Keep it enabled for reference-compatible processing.
- Disable it to use the extension’s earlier compact processing.
- Older saved workflows receive the enabled default. Image-block position calculations were also corrected to use whole-number positions. Some older seed-and-size combinations therefore produce slightly different images now, independently of the padding setting.
Prompt rewriting
The HunyuanImage 3.0 Prompt Rewriting node lets the model expand a short instruction before generating.
For example, a brief idea such as “a cozy reading corner” can be expanded into a description covering furniture, lighting, materials, and composition.
To use it:
- Use Instruct or Instruct-Distil.
- Make sure the matching
cot_headfile is available. - Connect the node to the encoder’s
prompt_rewritinginput. - Set
enabledtotrue. - Read the expanded text from the encoder’s
rewritten_promptoutput. The supplied workflows include rewriting, but leave it disabled.
Original prompts compared with rewritten prompts and reasoning-first rewriting
| Mode | What happens | Practical consideration | |---|---|---| | rewrite only | Generates an expanded description | Default; commonly a few hundred tokens | | think + rewrite | Generates a reasoning stage before the expanded description | Often exceeds 1,000 tokens and takes considerably longer |
Rewriting is much slower than ordinary prompt encoding. The author measured approximately 0.8–0.9 seconds per token, so a roughly 300-token rewrite can take four to five minutes.
For editing, the rewrite also considers your input images. Rewriting has its own seed, fixed by default, so changing only the image seed does not require a new rewrite.
Spectrum acceleration
Spectrum skips some expensive model calculations and predicts their intermediate results from earlier steps. Think of it as estimating selected in-between stages instead of fully calculating every one.
To enable it, place HunyuanImage-3.0 Spectrum between the model loader and KSampler, then set enabled to true. The supplied workflows already include it, disabled.
Spectrum disabled compared with Spectrum enabled
| Model | Warm generation without Spectrum | With Spectrum | Reported speed-up | |---|---:|---:|---:| | Instruct-Distil, 8 steps | 26 seconds | 20 seconds | 1.3× | | Instruct, 50 steps | 5 minutes 8 seconds | 1 minute 30 seconds | 3.4× | | Base, 50 steps | 4 minutes 58 seconds | 1 minute 27 seconds | 3.4× |
Spectrum is most useful on the 50-step models. It can change fine details and sometimes composition, so disable it when you want the full unaccelerated render.
The flex_window setting controls the speed-versus-fidelity trade-off. Larger values generally skip more full calculations, increasing speed while allowing more divergence. Start with the supplied defaults.
ComfyUI-HunyuanImage3 Performance
The following measurements come from the author’s RTX 4090, using PCIe 4.0 ×8, 188 GB system RAM, and 1024 × 1024 output.
| Model | Format | First image, cold | Later images, warm | |---|---|---:|---:| | Instruct-Distil | W4A8 | 39 seconds | 26 seconds | | Instruct-Distil | int8 ConvRot | 1 minute 10 seconds | 49 seconds | | Instruct-Distil | bf16 | 12 minutes 36 seconds | 12 minutes 6 seconds | | Instruct | W4A8 | 5 minutes 23 seconds | 5 minutes 8 seconds | | Instruct | int8 ConvRot | 9 minutes 48 seconds | 9 minutes 50 seconds | | Base | W4A8 | 5 minutes 14 seconds | 4 minutes 58 seconds | | Base | int8 ConvRot | 10 minutes 13 seconds | 9 minutes 50 seconds |
These are examples, not guaranteed timings.
- Cold includes loading and initial GPU transfers. A first run after reboot may take longer because the model is not yet cached from disk.
- Warm means the model is already in memory.
- Instruct and Base make two model passes per step; Instruct-Distil makes one.
- Ordinary prompt encoding is negligible in the reported tests. Prompt rewriting is a separate, expensive operation.
- VAE decoding took approximately 1.2 seconds.
- The bf16 test ran out of available RAM and had to stream from an NVMe drive, greatly increasing render time.
Smaller GPU results
The author tested reduced VRAM availability on the same RTX 4090:
| Available VRAM | First image | Later images | |---|---:|---:| | 24 GB | 39 seconds | 26 seconds | | 16 GB | 47 seconds | 30 seconds | | 12 GB | 47 seconds | 32 seconds |
These tests used Instruct-Distil W4A8 at 1024 × 1024 and produced identical images.
These are VRAM-limit tests, not benchmarks of separate 12 GB or 16 GB graphics cards. A slower GPU or narrower PCIe connection can take longer.
Changing attention backends is unlikely to provide a major speed improvement here: weight transfers dominate the workload. The author found similar per-step speeds across tested backends, although fine image details could change.
ComfyUI-HunyuanImage3 Large Images and Experimental Tools
Understand the resolution limits
Custom sizes work up to approximately four times the native image area, such as 2048 × 2048. However, generating a larger image from scratch tends to widen or expand the scene, rather than simply enlarge the same composition. Beyond the supported area—such as 2304 × 2304 and above—the author reports noise output. For larger artwork, the safer route is to generate at a native size and upscale.
Experimental nodes
These tools are under development. Their inputs, defaults, and behavior may change, and the supplied workflows do not use them yet. Test one tool at a time.
These nodes sit between the model loader and KSampler.
| Node | Artist-facing purpose | Reported trade-off | |---|---|---| | SEGA | Adjusts position scaling and attention behavior for larger images | No extra time reported | | SPA | Helps retain a native-size-style composition with larger subjects during early steps | About 30% additional time; default is three affected steps | | DyPE | Changes position scaling progressively during sampling | No extra time reported; text usually worked better without it | | HAP | Restricts attention to nearby image regions after the first step | About 12% faster per step at the tested 2× size | | Looped Diffusion | Repeats selected model layers for extra processing | Approximately 15–27% more time per step | | Tiled Refine | Refines large upscaled images in overlapping sections | Time increases with the number of tiles |
For Looped Diffusion:
early x2repeats layers 4–9 and can sharpen edges, add detail, and make text denser.mid x2andmid x2 halfrepeat layers 11–20 for counting-related experiments, with a risk of losing subjects.
Tested guidance
The author’s limited test sweeps found:
-
At 2× width and height, SEGA, SPA, or Looped Diffusion
early x2could each repair some large-image composition failures. -
SPA + Looped Diffusion
early x2produced colored speckles. -
SEGA + SPA also produced speckles at 2× per side.
-
A conventional upscale-and-refine pass worked up to 2× with
denoisearound 0.3–0.5. -
At
denoise: 0.7, faces began changing. -
At 3×, a conventional refinement pass introduced faint grid textures. For refinement beyond 2×, the author recommends trying Tiled Refine with:
-
A native tile size, such as 896 × 1152.
-
256-pixel overlap.
-
Prompt encoding at the tile dimensions.
-
A short, general description of the image.
-
denoise: 0.3. Keep Spectrum and Looped Diffusion disabled during tiled refinement. In the reported tests, they introduced mottled textures or speckles.
ComfyUI-HunyuanImage3 Node Reference
| Node | Purpose | |---|---| | Load HunyuanImage 3.0 Model | Loads a checkpoint and automatically identifies the model | | HunyuanImage 3.0 Text Encode | Prepares text-to-image conditioning | | HunyuanImage 3.0 Image Encode | Prepares an editing prompt and one to three reference images | | HunyuanImage 3.0 Prompt Rewriting | Enables prompt expansion before generation | | HunyuanImage 3.0 Guidance | Sets Instruct-Distil’s built-in guidance strength | | HunyuanImage 3.0 Resolutions | Provides the 37 native image sizes | | HunyuanImage 3.0 Empty Latent | Creates the starting latent with the correct shape | | Load HunyuanImage-3.0 VAE | Loads the shared VAE correctly; the standard loader misidentifies it | | HunyuanImage-3.0 Spectrum | Enables optional accelerated sampling | | HunyuanImage-3.0 SEGA / SPA / DyPE / HAP | Experimental larger-image controls | | HunyuanImage-3.0 Looped Diffusion | Experimental repeated-layer processing | | HunyuanImage-3.0 Tiled Refine | Experimental overlapping-tile refinement |
The remaining workflow components are standard ComfyUI nodes, including KSampler, VAE Decode, Save Image, and Select Model Device for multi-GPU placement.
What’s New with ComfyUI-HunyuanImage3
Extension changes documented by the author
The supplied documentation does not include a numbered extension changelog. It does describe these changes from earlier releases:
| Change | Why it matters |
|---|---|
| Redesigned nodes | Image Encode and Prompt Rewriting are separate nodes, and the old model_type setting is removed. Older workflows may need replacement. |
| Shared VAE file | All three models now use hunyuan_image_3_vae_fp16.safetensors. |
| Reference vision padding | Enabled by default to match Tencent’s reference image processing. |
| Corrected image-block position calculations | Some older images render slightly differently, especially with affected reference-image grid shapes. |
| Experimental large-image and looping nodes | Add options for composition repair, detail, acceleration, and tiled refinement, but require careful testing. |
Upstream model milestones
These dates refer to Tencent’s model project, not numbered releases of this ComfyUI extension:
- January 26, 2026: Instruct and Instruct-Distil were released, adding instruction-based editing, built-in prompt enhancement, and an eight-step distilled option.
- October 30, 2025: Tencent announced vLLM acceleration for its reference implementation. This is separate from the extension’s native ComfyUI acceleration.
- September 28, 2025: The original model, inference code, and technical report became publicly available. Multi-turn interaction remains listed as an upstream planned feature; it is not documented as available in this extension.
Troubleshooting ComfyUI-HunyuanImage3
Images look wrong, noisy, or unstable
- Return to the supplied workflow for your model.
- Select
eulerwith thesimplescheduler. - Restore the recommended steps and CFG.
- For Instruct-Distil, use KSampler CFG
1.0and Guidance2.5. - Choose a native resolution.
- Disable experimental nodes while checking the basic workflow.
An older workflow shows red nodes
- Open a current workflow from
workflows/. - Re-select your model and supporting files.
- Transfer your prompt and reference images into that workflow. The first-release node design is no longer the current layout.
The VAE is missing or loads incorrectly
- Check for
hunyuan_image_3_vae_fp16.safetensors. - Confirm that it is in
ComfyUI/models/vae/. - Use Load HunyuanImage-3.0 VAE, not the standard VAE loader.
A log message about missing
temb_projkeys during loading is documented as harmless.
Generation is very slow or runs out of memory
- Close other GPU-heavy applications.
- Use W4A8, the lightest supported weight format.
- Check available system RAM, not just VRAM.
- Use Instruct-Distil for eight-step sampling.
- Disable prompt rewriting to measure image-generation speed separately.
- Compare a later, warm image rather than only the first render. If the model cannot remain in RAM, disk paging can make generation dramatically slower.
Prompt rewriting produces no text
- Confirm that you are using Instruct or Instruct-Distil, not Base.
- Check that rewriting is enabled.
- Increase
max_new_tokensto 512–768 if the text-generation budget is running out. - Try rewrite only before using the longer think + rewrite mode.
A message says prompt rewriting needs a text head
- Open your chosen model’s ComfyUI weight repository.
- Obtain the matching
cot_headfile. - Place it in
ComfyUI/models/diffusion_models/. - Use
cot_head: autoto let the node find the appropriate file.
An edit is oversaturated, overprocessed, or only color noise
- Change the editing seed.
- Avoid reusing the seed that generated the source image.
- Try another seed-and-size combination if the problem remains. The author identifies this as a known issue under investigation.
Flash Attention repeatedly falls back to SDPA
This warning is harmless. The model needs an attention mask that the tested Flash backend cannot handle.
If your ComfyUI launch configuration includes --use-flash-attention, remove that flag to silence the repeated warning. Switching attention backends is not expected to solve the model’s main transfer-speed bottleneck.
ModelSamplingSD3 shift has no effect with Spectrum
A shift applied before Spectrum is currently lost. This is a documented known issue. Disable Spectrum when testing the shift on its own.
Large-image experiments produce speckles or mottling
- Disable all experimental additions and confirm the basic render works.
- Re-enable only one composition tool at a time.
- Avoid SPA combined with Looped Diffusion
early x2. - Avoid SEGA combined with SPA at 2× per side.
- Keep Spectrum and Looped Diffusion off tiled passes.
Frequently asked questions
Which model should you try first? Instruct-Distil W4A8 with its supplied text-to-image workflow.
Does a 12 GB GPU mean a low-memory computer is enough? No. The smaller GPU still relies on approximately the same substantial system-RAM budget.
Will bf16 necessarily make better-looking artwork? No. It is the original numerical reference, but visual preference is different from similarity to a reference render.
Can you use a separate negative prompt? The supplied encoders generate the model’s empty negative conditioning automatically; the workflows do not provide a second negative-prompt text box.
Why does rewriting take longer than drawing? Rewriting streams the model’s experts repeatedly for individual text tokens. An eight-step image can finish much sooner than a detailed text rewrite.
Can you edit with Base? No. Use Instruct or Instruct-Distil for editing and multi-image fusion.
Can you generate a larger image instead of upscaling? Within the documented area limit, yes, but the composition can expand rather than remain an enlarged copy. Native-size generation followed by refinement is the safer route for large artwork.
Learn More about ComfyUI-HunyuanImage3
Visual examples and practical starting points
- Author’s comparison gallery — compare models, compression formats, editing, rewriting, and Spectrum visually.
- Example workflows — ready-made graphs with file notes and model-specific defaults.
- Extension repository — current documentation, code, and project issue discussions.
- Tencent prompt handbook (https://docs.qq.com/doc/DUVVadmhCdG9qRXBU) — additional prompt-writing guidance from the upstream project.
Model documentation
- Tencent HunyuanImage-3.0 repository — official examples, model background, and upstream project discussions.
- Base model card
- Instruct model card
- Instruct-Distil model card
- HunyuanImage-3.0 technical report — deeper architectural and evaluation details.
Spectrum background
- Spectrum project
- Spectrum visual project page
- Spectrum paper Spectrum’s published results for other image and video models are not performance guarantees for HunyuanImage-3.0. Use this extension’s measurements as the relevant reference.
ComfyUI-HunyuanImage3 Licensing
The extension code and model weights have different license terms:
- Extension code: GPL-3.0.
- Tencent-derived model code: remains subject to the Tencent Hunyuan Community License.
- Spectrum forecaster: adapted from the MIT-licensed Spectrum project.
- Model weights: governed by the Tencent Hunyuan Community License, including its Acceptable Use Policy. The author notes that the Tencent license does not apply in the European Union, the United Kingdom, or South Korea. Review the actual license terms and territorial restrictions before using the weights, particularly for professional work.
See NOTICE for additional attribution and licensing details.
