Sulphur 2 Review: Using the 9B LTX 2.3 Fine-Tune Online

Sulphur 2 is an open-source AI video generation model, distributed by the SulphurAI community on Hugging Face as SulphurAI/Sulphur-2-base. It was released on 2026-05-03 — at the time of writing, the model is barely three weeks old.

The technical lineage: Sulphur 2 is a 9-billion-parameter fine-tune of Lightricks' LTX 2.3, the 22B open-source video model released earlier in 2026. SulphurAI took LTX 2.3 as the base and trained additionally on roughly 125,000 video clips, with the training data filtered to exclude 2D and animation content in favor of realistic scenes. The release shipped with open weights, distill LoRAs, four official ComfyUI workflows (text-to-video and image-to-video, base and distilled variants), and a local prompt enhancer — a self-contained bundle designed for local deployment, not a hosted SaaS.

That last point is what creates the gap this site fills. Running Sulphur 2 locally needs roughly 24–32 GB VRAM for the base safetensors; community GGUF re-quants bring it down to 10–23 GB on disk, but still ask for serious hardware. Most creators searching for Sulphur 2 do not have that hardware lying around.

sulphur2.net is an independent online hosting service for that same open-source model. This site is not operated by SulphurAI; we run the model behind a browser interface with credit-based generation, a saved library, and the usual T2V / I2V modes. You type a prompt, optionally upload an image, choose duration and aspect ratio, spend credits, and get a clip back. The model behavior is Sulphur 2 — the difference is that we manage the GPU.

Honest framing first: Sulphur 2 has no published technical report, ablation study, or third-party benchmark at the time of writing. Quality claims rest on the official README, the SulphurAI release thread, and community comparisons that started appearing in mid-May 2026. The model is too new for formal evaluation, and any reviewer claiming a verified score for Sulphur 2 specifically is overreaching.

What can be said responsibly: the base model is well-measured. Lightricks' LTX 2.3 scores Elo 1121 on the Artificial Analysis open-weight video leaderboard (as LTX-2.3 Fast), which makes it the #1 open-source video model on that board. Sulphur 2 inherits the same base architecture and weights and adds a realism-focused fine-tune on top.

For closed-model context: the same Artificial Analysis board places Kling 3.0 at Elo 1243 (the current #1 across publicly accessible models), Runway Gen-4.5 at 1225, and Veo 3.1 at 1217. HappyHorse 1.0 scores higher still (T2V Elo 1361, I2V Elo 1398) but is not publicly accessible — its API has not opened.

Because Sulphur 2's added training (~125K clips) explicitly filtered out 2D and animation content, the model is likely tuned for realistic scenes, micro-expressions, and atmospheric continuity. That same focus is likely to weaken stylized, illustrated, or animation prompts compared with the base LTX 2.3.

For short product motion, image-to-video on a clean reference, and realistic-style social clips: expect output quality in the LTX 2.3 ballpark, possibly sharper for realistic faces and skin texture, and within the bounds of what a top open-source video model produces in mid-2026. For animated / illustrated / cartoon styles, results may be weaker than the LTX 2.3 base would have produced before fine-tuning. If your project is non-realistic, test the 50 free credits before committing.

If you want to see actual output, the Sulphur 2 Showcase gallery has clips organized by use case.

Sulphur 2 is known in the open-source community for having fewer content restrictions than several mainstream commercial AI video models. That positioning shows up in most community write-ups and is part of why the model attracted attention so quickly after release.

That property belongs to the open-weight model, not to this hosting service. sulphur2.net runs Sulphur 2 under its own Terms of Service, which apply to every generation regardless of the base model's permissive nature. Acceptable use, content categories, and commercial rights are governed by the terms of service and the refund policy — read them before running campaigns or client work through this site.

If you intend to use Sulphur 2 for content categories outside what sulphur2.net's Terms allow, the open weights are available on Hugging Face (SulphurAI/Sulphur-2-base) for local deployment.

The signup-to-first-clip economics are easy to describe. 50 free credits at signup. One 5-second 720P generation uses 50 credits, so the free tier covers exactly one test. Credits never expire — a buyer can purchase a pack and spend it across weeks or months. Library retention is 6 months: download any clip you need long-term; the library is for review and re-download, not permanent archive.

The full plan-by-plan pricing lives on the pricing page — confirm exact prices there. For external comparison: Kling 3.0's commercial entry plan starts at about $6.99/month according to public review coverage; Seedance 2.0 access varies by gateway and partner; HappyHorse 1.0 is not yet commercially available.

Practical credit math, for first-time buyers planning a real project: a usable creative direction typically takes three to four generations (one to establish, two to three to refine). Budget around that rather than around the single free test.

If your project fits one of the scenarios below, the 50 free credits are the right next move. Start with the generator and the prompt guide for the prompt patterns that work best with the model.

Sulphur 2 is narrow on purpose. For the projects below, a different model class will fit better — see the alternatives section after the failure modes.

Three honest patterns to expect, all consistent with the LTX 2.3 model class. Plus one pattern specific to Sulphur 2's fine-tune.

If Sulphur 2 doesn't fit the project at hand, the most relevant alternatives in May 2026 are below. Each comparison page covers the full spec-by-spec breakdown.

Sulphur 2 is the fastest way to use LTX 2.3 today. Output quality matches the open-source state of the art, the browser workflow removes the GPU and ComfyUI setup tax, and the credit model is straightforward. The honest limit is audio: if a project needs native generated sound in the same pass, this is not the tool for that step.

For short product clips, image-to-video animation, and silent social drafts — start with the 50 free credits and a single tested direction. If audio matters, run a parallel test on Kling 3.0 or wait for HappyHorse 1.0 to open access. If full creative control matters more than convenience, install LTX 2.3 locally.

Bottom line: worth a free test for anyone whose use case fits the four "Who Should Use" categories above; worth a credit pack only after the free test confirms the direction.