Sulphur 2 is a community fine-tune of Lightricks' LTX 2.3 that runs on your own GPU with no service-side prompt filter. Here is what the download contains, how much VRAM it really needs, how the ComfyUI workflows run, and where it beats or trails Runway and Kling AI.
Sulphur 2 is a community fine-tune built on Lightricks' LTX 2.3, the open-weights video model that renders a clip and its audio in one pass. The architecture is not new, but the weights are downloadable, they run on your own GPU, and text-to-video and image-to-video are both supported.
Most people arrive because of one word: uncensored. It means something narrower in an open-weights release than it sounds. This review answers four questions in order: what is in the download, what hardware runs it, how the ComfyUI workflows behave, and how it compares with Runway and Kling AI.
Sulphur 2 is not a new architecture. It is a community fine-tune of Lightricks' LTX 2.3, the 22B-parameter diffusion transformer that renders video together with synchronised audio. It trained on a curated set of roughly 125K clips weighted towards people and everyday motion, the part of video generation where base models usually wobble first. Facial consistency and body movement hold up better than on the base checkpoint; landscapes look much the same.
The download is a family of files, not one checkpoint: a bf16 dev weight near 16.7 GB, an fp8-mixed weight near 8.5 GB, a distilled variant and a step-cutting distilled LoRA. The model card puts the effective parameter count near 9B while the transformer underneath is the 22B LTX 2.3 stack, so read the card, not a summary post. Two extras matter as much as the weights: a GGUF prompt enhancer, and ComfyUI graphs for both tasks in base and distilled form.
Open weights are the mechanism. The model runs on your hardware, so no service-side filter stands between your prompt and the frames, and the release ships two prompt enhancers: the alternative variant drops the filtering step, so a request is not quietly rewritten into something tamer. That is the whole trick. Nothing in the weights makes the model more capable; it removes an intermediary.
Two caveats usually get skipped. The licence is a custom non-commercial and research licence, so open weights are not free to ship in paid client work. And once the filter is gone, moderation is your job: the same local run that storyboards a fight scene can produce material that is illegal where you live. That is a legal position, not a feature.
An 8.5 GB fp8 file does not mean a mid-range card is enough: the transformer is not the only resident component. LTX 2.3 pairs it with a Gemma-3-12B text encoder, a video VAE and an audio VAE. A published recipe for the full-precision model on a 96 GB card reports weights loading near 36 GB and inference peaking near 62 GB at 768x512 with 30 steps. On a 24 GB card the usual route is a quantised transformer with the text encoder in system memory.
| Variant | File size | Typical card | What changes in practice |
|---|---|---|---|
| bf16 dev | 16.7 GB | 24 GB and up | Full precision; the text encoder and VAEs compete for memory |
| fp8-mixed dev | 8.5 GB | 16 to 24 GB | The practical starting point on one consumer GPU |
| GGUF quantised builds | varies | 12 to 24 GB | Conversions trade motion detail for headroom |
| Distilled weight plus LoRA 1.1 | same as base | as above | Fewer sampling steps, so iteration is cheaper |
| Full-precision serving | 36 GB loaded | 80 to 96 GB | Reference path: peaks near 62 GB at 768x512 |
LTX 2.3 runs at 24 fps and needs resolutions divisible by 32; the published recipes use 768x512, with 512x384 as the smaller option. Duration is a frame count, not a seconds field: frames = seconds x 24 + 1, so 3 seconds is 81 frames, 10 seconds is 241 and 20 seconds is 481. Long clips are harder to keep coherent, not merely slower, so most usable work lands in the 3 to 10 second band.
| Aspect | Text-to-video | Image-to-video |
|---|---|---|
| Input | A prompt, optionally expanded by the enhancer | A still image plus a motion prompt |
| Strength | Framing, camera and lighting are yours | Identity and composition stay locked to the source |
| Typical failure | The subject drifts mid-clip; props lose consistency | The model under-animates, slides the frame, or invents unsupported motion |
| Good first use | Exploring a sequence you have not blocked out | Animating a still you like: a render, a photo, a character sheet |
A one-line prompt is an under-specified brief, and the sampler fills the gaps differently on every seed. The bundled enhancer is a small GGUF language model that also runs in LM Studio or any compatible interface, expanding an idea into a shot description with subject, wardrobe, camera move and lighting. Treat it as a drafting tool and keep the prompts that worked: prompt drift is the usual reason the same idea looks different a week later.
| Dimension | Sulphur 2 local | Runway | Kling AI |
|---|---|---|---|
| Runs on | Your GPU, weights on disk | Vendor cloud or API | Vendor cloud or API |
| Cost model | Hardware and electricity, no per-clip credit | Credits per generation | Credits, tiered by version and length |
| Content policy | No local filter; licence limits commercial use | Provider terms and filters apply | Provider terms and filters apply |
| Control | Resolution, frame count, sampler, seed, LoRA | Curated controls, preset camera moves | Curated controls plus preset camera moves |
| Best fit | High iteration volume, private material | Teams without a GPU wanting a managed API | Polished short shots where presets do the work |
This is not a ranking. Hosted services win the first afternoon: no driver install, no 30 GB download, no memory tuning, and a credit price you can quote. They lose on volume and control: a hundred iterations consume real credits, and the sampler settings behind a good frame may be hidden. A local fine-tune flips both, because the hundredth clip costs the same as the first and every knob is yours. Prices and terms change often on both sides, so check the vendor pages.
Open3DAI does not generate video, and a review implying otherwise would be useless. It generates 3D assets: text-to-3D, image-to-3D, multiview input, AI texturing, and side-by-side engine comparison before export. Export covers GLB, glTF, OBJ, STL, PLY and 3MF, plus 48 converters, 12 browser tools and a REST API. This deployment runs one local procedural engine plus optional third-party adapters you connect with your own API keys; it does not host or train other vendors' models. The bridge between them is frames: turntable renders go into image-to-video, and a good frame from a clip becomes an image-to-3D reference.
Engine availability, credit pricing, licence terms and quantisation formats change between releases, and every figure here is what the vendors published at the time of writing. Verify on the vendor page before committing a budget.
A fine-tune. It is trained on top of Lightricks' LTX 2.3, so it inherits that architecture, its 24 fps output and its divisible-by-32 resolution rule, and it adds its own training set, prompt enhancers and ComfyUI graphs.
It means the weights run on your machine with no service-side prompt filter, and one of the two bundled prompt enhancers skips the filtering step. It does not change what the model can render, and it does not exempt you from the licence or from local law.
The fp8-mixed weight is about 8.5 GB, but the text encoder and the VAEs need memory too. A 24 GB card is the practical floor with quantisation and CPU offloading; the published full-precision serving recipe targets a 96 GB card, reports roughly 36 GB loaded and peaks near 62 GB at 768x512.
Not by default. The release carries a custom non-commercial and research licence, so commercial use is something you have to clear with the rights holder rather than something open weights grant automatically. Read the terms on the model page first.
Duration is a frame count at 24 fps, and the published recipes demonstrate 81, 241 and 481 frames, roughly 3, 10 and 20 seconds at 768x512. Coherence usually degrades before the runtime limit does, so 3 to 10 seconds is the useful band for most shots.
Choose hosted services when you need a result today, a managed API and pricing you can quote per clip. Choose the local fine-tune when you need many iterations, private material, sampler-level control and no per-clip cost. Plenty of teams run both and route work by shot type.
No, but it is the shortest path to a first render because the release ships wired graphs. The weights are standard safetensors and GGUF files, so any interface that supports LTX 2.3 pipelines can load them, and the prompt enhancer runs in any GGUF-compatible UI.
No. Open3DAI generates 3D assets: text-to-3D, image-to-3D, multiview input and AI texturing, with export to GLB, glTF, OBJ, STL, PLY and 3MF. Video generation is out of scope, so run a video model next to it and pass frames between the two.
Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.
Flux 3 is Black Forest Labs' third generation model and the first in the line to render video and audio next to images. This review covers what Self-Flow changes, what the published material does and does not say, how it stands beside Runway and Sora, and what access costs.
MiniMax H3 Max is fal's post-trained version of the open-weight MiniMax H3 video model, tuned for prompt adherence and served on an inference stack co-designed with it. Here is what it changes, how to read the 35x throughput figure, what a second of output costs, and when the stock H3 endpoint is still the right pick.
An editorial ranking of ten hosted services and open-weight engines, with the criteria stated up front, two comparison tables, and a plain account of the cleanup every generated mesh still needs.