radiance
32-bit Ai Generation with Color Science · GPU Accelerated
Nodes (156)
A spec-check node that tells you the truth about your transform
Still works, still hidden, probably not the one you want
The one ACES node in this pack you actually want
Stop your neon from turning into flat pink
The curve that decides what 'correct' looks like
Find, download and install the OCIO config without leaving ComfyUI
Put your colour pipeline in a sidecar Resolve can read
The cheap ACES look for previews (and why it isn't a deliverable)
A real upscaler collection — with the HDR mode most upscalers are missing
Make it look shot on glass, not generated
Replay the exact grade you saved — on any shot, any time
Getting Alexa 35 footage into your ACES working space
Put your cuts on the beat without opening a DAW
Subtitles out of ComfyUI, no separate app required
Going from float to 8-bit without the banding
See the banding before your client does
Eight blend modes that know what HDR is
One JSON file, and every optical node agrees on where the camera was
Write the grade to a file so it can leave ComfyUI
Read a .cdl file and turn it into ten numbers you can wire anywhere
The grade that survives being handed to someone else
The two-pixel lie that makes an image look shot, not rendered
Why naming the actual glass beats 'cinematic, 8k, masterpiece'
A prompt builder that knows whether you're on Flux or SDXL
Find the highlights your SDR image already threw away
Getting ComfyUI out of sRGB jail
The node that writes down what your pixels are
Deliberately making your images look compressed
Review 48 frames in one glance instead of 48 clicks
A ControlNet apply node with a Union mode picker built in
The apply node for Radiance's sampler, union models included
Real curves, typed as JSON, that don't clip your HDR
Get your frame out of ComfyUI and into the timeline
One node for Blackmagic's colour space
A real multiscale denoiser, with the honest caveats in the tooltips
Depth Anything V2 in your graph, with the licence you didn't read
Cinematic defocus from a depth map, bokeh shape and all
Read footage like a VFX shot, not a web image
The save node that speaks EXR and ProRes
Downscale without destroying the float range
Local CFG, without stacking three samplers to fake it
Beauty, depth, normals and albedo in one file the comp can open
The whole pass bundle, with guard rails against lossy data
The zone read-out that tells you what's actually clipped
Film grain that doesn't look like a noise filter
The fast, ugly, shareable render preview
Exposure, contrast and lift/gain that don't clip your HDR
Get your image into float and linear without losing the top end
See which edges are actually sharp
Because metadata does not survive a client review
One node for exposure, gamma, clamp and normalize
Radiance Grade — a real Lift/Gamma/Gain node so you stop faking color work with img2img
Stop grading the preview and grade the actual pixels
Radiance Grade Match — color match your AI frame to a reference in one step
Turn one image into an environment map your 3D app will load
Is your output actually HDR, or just bright?
Let the histogram pick your camera log
Proof that your HDR blend did anything at all
The primaries swap you'll use twenty times a day
Linearise, adapt, convert, compress — in one node
Fix one eye at 4K without regenerating the face
The node that tells you your compression ratio was too hot
Scene-linear in, a signal your TV will actually read
Stretching an SDR image into HDR, honestly
You shot the brackets, now merge them properly
Grain that scales with exposure instead of eating your highlights
Only the blown-out parts come from the model
What it did, and what to use instead
Patch the model without torching your cached checkpoint
It reads the metadata your HDR LoRA is carrying
See what you're actually making on the display you actually own
Real colour management, minus the config wrangling
Undo the normalisation without re-deriving the math
The 20-line fix for banded HDR latents
Pull detail back without clipping anything
Putting a fixed patch back without a visible seam
Real headroom on top of an SDR image, and the limits of that
Nine presets that overwrite each other, and how to actually drive it
Why you shouldn't wire this one up
Getting real scene-linear values out of a latent
The encode that survives above 1.0
Radiance Highlight Synthesis — fake the roll-off that flat white skies deserve
Hue-vs-hue curves, except you type the curve as JSON
One reference image in, video latent out — plus the widgets that only rewrite your prompt
Radiance's \"Read Image\" node loads models, not images — here's what it's actually for
Barrel, pincushion, and an ST map you can hand to Nuke
A smoke alarm for wrongly-tagged colour, not a truth detector
Turning a rough mask into an edge-following alpha, with no model
Radiance Load Image — a loader that keeps your mask separate from your pixels
Radiance LoRA Stack — five LoRAs in one node so your graph stops looking like a fire escape
'this window should be 1000 nits', as a node
Radiance LUT Apply — drop a real .cube LUT into ComfyUI and stop hand-tuning grades
Radiance LUT Blend — stop choosing between looks, blend two of them
EXR and video out to Nuke, Resolve or Fusion — and a TCP server you probably shouldn't open
Radiance Motion Blur — the post pass that makes AI video stop looking like stop-motion
Choosing between candidate masks without re-running the segmenter
Real renderer passes from a multilayer EXR, not guesses
An alpha-over that knows about depth, shadows and light wrap
Pull a full VFX pass set out of a flat image, with real models and a real download
Multipass Extract is retired — here's what replaced it
Move the key light in comp instead of re-rendering
Stream your ComfyUI render straight into OBS, Resolume or Nuke
Get ComfyUI output into Nuke without the path arguments
Load your OCIO config once, then stop worrying about colour
Stop guessing colour-space names — list them instead
Dense optical flow without the OpenCV homework
Turn delivery specs into a pass/fail you can automate
Watch a render from another machine's browser
Version your graph like a shot, not a screenshot
Radiance Pro Upscale — clean vectorized upscaling when you don't want a model rewriting your image
The node that tells you your delivery is technically broken
EXR, DPX, ProRes and sequences — the loader ComfyUI should have had
Two characters, four prompts, one canvas — the grid way
Boxes, masks and prompts that stay where you put them
One synthetic light, one normal map, no re-render
One dropdown that sets your resolution, latent channels, and even video length
Add the skew, wobble, and flash bands that make a still read as phone footage
This SAM keyer ships empty — here's what to use instead
Why it loads a model into nothing
The sampler that replaces your KSampler, CFG++ hacks, and video samplers in one node
Find the cuts without scrubbing the timeline
Pull one shot out of a clip, by index
Fake the headroom, don't expect magic
The pre-VAE route, and whether you still need it
The node that would rather fail than fake it
SDR → HDR Universal is the Radiance node you'll actually use
One JSON table, four grade numbers per shot
How to actually see what your grade changed
Translation only, and it hands you the exact pixel offset
A whole text-to-video graph in one node, minus the honesty
Stop a mask from popping before you stitch
One loader for Flux, SDXL, WAN, HunyuanVideo — and it picks the architecture itself
Upscale to an exact target size without cropping, stretching, or blowing up VRAM
The pass you run before you look at the eyes
Three model tiers, one dropdown, and no more upscaler shopping
It measures noise and sharpness, and it does not know faces
The tiling engine, plus the fix for upscaler colour drift
Windowed, not temporal — and why that distinction costs you
A stateful buffer living inside a stateless engine
The VAE decode with the one knob you actually need
Three conditionings, three very different merge modes
Writes EXR and GIF, and not one movie codec
Pull one frame out and put it back
It conditions a video model, not your pixels
You feed it pixels, not latents
The correctly-shaped noise your video sampler keeps asking for
One loader for LTX, Wan and HunyuanVideo, and the Wan 2.2 dual-expert trick done for you
Rotoscope one frame, let the flow do the other twenty-four
The node that stops your video sampler from guessing
Stop retyping the same HDR words into every clip
The KSampler-shaped part of a video graph, minus the shape guesswork
A VFX review tool in a node graph, and this one doesn't dead-end
The cheapest trick for making an AI render look shot rather than generated
Fix a cast with a Kelvin number, not a guess
The node that treats your whole workflow as a saveable project
Get your frames off the browser canvas and onto the NAS, in whatever format the job wants
<img src="r_icon.png" width="76" alt="Radiance mark"><br> <img src="RADIANCE.png" width="640" alt="Radiance">
HDR, colour management, VFX tools, and image review inside ComfyUI.
Radiance brings 147 visible nodes into your workflows for processing images and video, working with EXR and ACES, reviewing results, and delivering files to Nuke or DaVinci Resolve.
Install · Quick start · Models · Example workflows · Documentation · Support
</div>What you can do
| Your task | Radiance tools | |---|---| | Create and process HDR images | HDR VAE Encode and Decode, SDR → HDR tools, HDR diagnostics | | Manage colour | ACES, OpenColorIO, colour-space conversion, LUTs, and CDL | | Review images and video | Viewer with Simple and Advanced modes, scopes, comparison, and playback | | Prepare VFX shots | Masks, depth, optical effects, estimated passes, and relighting | | Save and deliver results | Read and Write nodes for images, video, and EXR sequences | | Organize your work | Project Manager, Workflow Library, Assets, and reusable Gizmos |
<div align="center"> <img src="viewer.png" width="920" alt="Radiance Viewer with image review, colour controls, and scopes"> </div>Installation
Install with ComfyUI Manager
- Open ComfyUI Manager.
- Search for Radiance and install it.
- Restart ComfyUI.
- Refresh your browser.
You can also find Radiance in the ComfyUI Registry.
Install manually
Run these commands using the Python environment that runs ComfyUI:
cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt
Restart ComfyUI and refresh your browser.
Windows portable users: use ComfyUI's bundled python_embeded\python.exe when installing dependencies.
Update Radiance
Use Update in ComfyUI Manager.
For a Git installation, run these commands inside the radiance folder:
git pull
python -m pip install -r requirements.txt
Restart ComfyUI and refresh your browser after updating.
Upgrading from an older version? Back up your workflows and check the migration notes. Some nodes and widget layouts changed in v3.5.
Quick start
Process an existing image
Start with this simple workflow:
Read → SDR → HDR Universal → Viewer
- Load your image with Read.
- Connect it to SDR → HDR Universal.
- Connect the result to Viewer.
- Run the workflow.
- Add Write to save the result as EXR.
The SDR-to-HDR model downloads on first use when downloads are enabled.
Generate an image
Open the included starter workflow in ComfyUI.
Select the FLUX.2 Klein 4B distilled model, FLUX.2 VAE, and Qwen3 4B text encoder in the Loader, enter your prompt, and run the workflow. Review the result in the Viewer and connect a Write node when you want to save it.
The required model files are separate downloads. See Models and downloads below.
<div align="center"> <img src="basic_workflow.png" width="920" alt="Radiance starter workflow from model loading to HDR review"> </div>Models and downloads
RUDRA for SDR to HDR
RUDRA on Hugging Face provides the learned SDR-to-HDR model used by SDR → HDR Universal and SDR → HDR Recover. The default model downloads on first use when downloads are enabled.
For manual or offline installation:
| Model | Download | Save in |
|---|---|---|
| sdr2hdr_shadow_v1.safetensors | Download default model | ComfyUI/models/radiance/ |
| Alternative RUDRA image checkpoints | Browse available models | ComfyUI/models/radiance/ |
Create the radiance folder if needed. Leave pixel_checkpoint empty to use the default model. To prevent automatic model downloads, set RADIANCE_ALLOW_DOWNLOADS=0 before starting ComfyUI.
Model licence: RUDRA weights are licensed for non-commercial use. Read the RUDRA weights licence before use. This is separate from Radiance's GPL-3.0 code licence.
Models for the starter workflow
The included starter workflow uses FLUX.2 Klein 4B distilled, with 4 sampling steps at 1024 × 1024. Download these three files, place them in the folders below, and select them in the Loader:
| File | Hugging Face download | Save in |
|---|---|---|
| flux-2-klein-4b-fp8.safetensors | FLUX.2 Klein 4B distilled | ComfyUI/models/diffusion_models/ |
| flux2-vae.safetensors | FLUX.2 VAE | ComfyUI/models/vae/ |
| qwen_3_4b.safetensors | Qwen3 4B text encoder | ComfyUI/models/text_encoders/ |
Use the distilled model listed above, not the base variant, which needs different sampling settings. Update ComfyUI for FLUX.2 Klein support and review each model's licence. See the official ComfyUI guide for more details.
Example workflows
Choose a workflow that matches your task:
| Workflow | Purpose | |---|---| | Starter workflow | Generate and review an image | | SDR to HDR | Convert an existing image to HDR and save EXR | | Colour grading | Adjust colour and save a graded image | | HDR VAE | Encode, decode, and compare an HDR image | | Multipass relighting | Estimate passes and relight an image | | HDR delivery | Create EXR, HDR10, and SDR outputs |
Drag a workflow JSON file onto the ComfyUI canvas to open it.
Using the Viewer
Start in Simple mode for image review, comparison, and playback. Switch to Advanced mode for scopes, colour controls, and detailed inspection.
- Hover or select the Viewer to use its keyboard shortcuts.
- VIDEO inputs can include audio. IMAGE batches do not contain audio.
- Use Write nodes for video and sequence export.
- Viewer adjustments affect the preview unless you apply equivalent processing in the workflow.
Before you start
- Additional models: some tools download models on first use. Multipass Estimate requires substantially larger downloads than SDR → HDR.
- Long-video windowing: experimental and disabled by default. Some scenes can distort when split into temporal windows.
- Viewer backend: WebGL is the default. Optional WebGPU support has feature limitations.
- Estimated passes: depth and material predictions are estimates, so results can vary and video may flicker.
- Temporary previews: rerun the workflow if ComfyUI's temporary preview files have been removed.
Read the full known limitations.
Documentation
In ComfyUI, hover over a node or input to read its description.
Support
If something does not work:
- Restart ComfyUI and refresh your browser.
- Check the ComfyUI console for errors or missing dependencies.
- Confirm dependencies were installed into ComfyUI's Python environment.
Report an issue with your Radiance version, ComfyUI version, operating system, error message, and a minimal workflow that reproduces the problem.
License
Radiance is licensed under GPL-3.0.
Downloaded model weights have their own licences. Review those terms before using or distributing them.
Built by FXTD Studios.