LTX-2.5 SDR-to-HDR LoRA: ACEScg EXR and HLG, not an MP4 API
LTX-2.5 SDR-To-HDR outputs scene-linear ACEScg EXR frames plus an HLG master from an SDR clip. Sume's video routes return MP4. Where each fits.

LTX-2.5 SDR-To-HDR is an open adapter that converts a standard-dynamic-range clip into HDR and writes scene-linear ACEScg EXR frames alongside an HLG master. It is a local pipeline, not an API. Sume's video routes return MP4 files, so if your finishing step needs EXR frames, this is a route Sume does not replace.
Facts below come from the SDR-To-HDR model card, read on 2026-10-02. The Sume side comes from Video generation and Video filter.
What does the SDR-To-HDR adapter take and return?
Per the card, the adapter expands highlight headroom and recovers range that an SDR grade compressed out. The SDR clip is itself the control signal. You can pass an SDR MP4 or MOV, or a directory of EXR frames, and choose the input colour space: sRGB gamma, sRGB, ACEScg or ACEScct. Outputs are ACEScg EXR frames and an HLG master.
- No text prompt: the card says it uses precomputed scene embeddings.
- Fixed at 8 distilled Euler steps.
- Frame count must be 8k+1.
- Needs the distilled transformer and video VAE from LTX-2.5, run through a dedicated
hdr_ic_lorapipeline rather than the standard video-to-video one. - ComfyUI is supported through the ComfyUI-LTXVideo extension.
Why does the EXR output matter?
EXR frames in scene-linear ACEScg are what a colour pipeline or compositing tool expects when it grades or merges plates. An 8-bit MP4 cannot carry that range. So the adapter sits at the end of a pipeline, after you have a finished SDR clip, and before the grade.
| Route | What you get back |
|---|---|
| LTX-2.5 SDR-To-HDR (local) | ACEScg EXR frames and an HLG master |
| Sume Video Router and /v1/videos | A video job whose result is an MP4 download |
| Sume Video filter | A new MP4 after dim, crop or an allowlisted filtergraph |
What can Sume do with the same clip?
A Sume video job completes with unsigned_urls pointing at a downloadable file, as the video docs describe. The docs do not describe HDR, EXR or a bit-depth option. Video filter is a pixel pass over one workspace clip that returns a new MP4 and does not describe a tone-mapping or colour-space conversion.
So the practical split is: generate or edit on Sume, download the MP4, then run SDR-To-HDR locally if you need HDR masters. Expect the adapter to invent highlight detail, because it is a generative model; it is not a lossless inverse of the grade. The card does not claim otherwise, and you should check the result against the intended look before delivery.
What does it cost to run yourself?
The card gives no price, and VRAM numbers are only described as 'sufficient' with memory limits for long or high-resolution content. Plan for your own GPU time and test one short clip first. The license is the LTX-2.x Community License, so read its terms for your company before you use outputs commercially.
Related reading on the Sume side: the earlier HDR and EXR export post explains why Sume stays MP4-only, and the frame count post explains the 8k+1 rule you hit here.
Sources
Related posts
More in Media tools
- Luma Photon image reframe: 10 MB source, 2432x1024 max vs Sume
Luma's Reframe API extends an image to a new aspect ratio with photon-1 (10 MB in, 2432x1024 out at most). Sume has no reframe call; here is the closest path.
- Mastodon video upload: 99 MB, one video, transcoded to 1300 kbps
Mastodon takes one video per post, up to 99 MB, and transcodes it to H.264 MP4 at up to 1300 kbps. What to send from Sume so the re-encode costs less quality.
- MiniMax H3 reference formats: HEIC, MOV, WAV, MP3 limits
MiniMax H3 accepts HEIC/HEIF images, MP4/MOV video and WAV/MP3 audio as references. A pre-flight checklist before you submit minimax-h3 on Sume.
- Keep the sound of AI clips in a Sume timeline: detach to a spine
A Timeline 1.0 render takes one audio spine. To keep audio generated with your clips, detach each clip's track, join the parts, and use that as the spine.
Written by Sume