Two community branches of LTX 2.3, both trained rather than merged, both aimed at the same problem: the base model is conservative. Sulphur 2 was retrained on a large corpus for more open generation; 10Eros is specialised for image-to-video and the author explicitly does not recommend it for text-to-video. Four official example graphs ship in the mirror, two per model.
Evidence note: this episode ships no workflow JSON, and 说明.txt records neither a workflow link nor a mirror URL. Everything below comes from the narration; no node count, sampler value or model filename is claimed.
When this is useful
- Material the base LTX 2.3 refuses to generate because of content restriction rather than capability.
- Image-to-video where you want more latitude than the distilled default allows.
- Audio-driven digital human, using the 10Eros graph that accepts an audio track.
- Getting to 30 fps output from a model that samples at 24, via the author's node.
How this differs from the previous build
What separates these from a merged checkpoint is that they were actually trained.
- Sulphur 2 is built on LTX 2.3 with heavy additional training, not a weight blend. The author's framing is freedom: it accepts prompts the base model will not.
- 10Eros is an image-to-video specialist. The author repeats the model author's own warning: do not use it for text-to-video.
- Two of the four graphs are audio-driven. One 10Eros graph takes image plus audio and behaves like a digital human; the other takes image only.
- The Sulphur graphs are two-stage: base generation, then an upscale model. That is close to the official example rather than the author's usual three-tier build.
- A custom node handles frame rate. It takes 24 fps sampling and produces clean 30 fps output; in these demos the author stays at 24 and does not enable the bump.
Running it
- Deploy the mirror and open ComfyUI; four graphs are in the left workflow list, top two 10Eros, bottom two Sulphur.
- Choose the 10Eros audio-driven graph and upload audio, then set the trim start and length — 20 seconds starting at 6.9 seconds, in the demo.
- Load the image. This is mandatory for every one of the four graphs.
- Set frame rate, prompt and negative prompt, then run.
- For the image-only 10Eros graph, drop the audio input; everything else stays the same.
- For Sulphur, set size, frame rate, frame count and prompts. The second stage is the upscaler.
- To load a LoRA, upload it to
ComfyUI/models/LoRAthrough the file manager, then add it at the LoRA node — Sulphur exposes one directly, 10Eros shares the same upload path.
Honest limits
- The author declines to demonstrate results in detail for these two models, so there is little visual evidence in the episode itself.
- 10Eros is image-to-video only by its author's own statement; text-to-video use is not covered.
- The 30 fps path is described but not exercised in the demo — the graphs run at 24 fps.
- No workflow link and no mirror URL are recorded in 说明.txt for this episode, and no Bilibili release exists.
- Nothing here is parameter-verified: there is no JSON to read.
Episode resources
- Video tutorial (YouTube): https://www.youtube.com/watch?v=8Zs0L8Oh-3Y
- Chinese original (Bilibili): not published for this episode
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)