> ⚠️ **18+ / NSFW.** This page contains adult examples and descriptions. If you are under 18, or do not wish to see explicit material, go back now.
> This is the uncensored continuation of [Betting on the video niche the big labs walked away from](/en/articles/residual-video-domain-i2v/) (the SFW version). Reading that context first is recommended.

## About this page

In the SFW version I kept things vague — "community fine-tunes," "uncensored models." Here I name them, give the real recipe, and show the examples. My own platform isn't bound by a content policy, so I can share the **actual state of the art** in a space the frontier models structurally avoid. That's the whole point of the experiment.

## The models I used

| model | type | take |
|---|---|---|
| LTX-2.3 22B (base) | official dev checkpoint | strong physical coherence; baseline for general quality |
| **Sulphur-2-base** | community uncensored fine-tune (T2V+I2V) | **clearly strong at the NSFW boundary — the one unambiguous differentiator.** Trades generalization, so anatomy can break under some conditions |
| **10Eros** | layer-scaled merge of the Sulphur line (I2V-focused) | used as an I2V starting point; motion character differs |

Bluntly: on general quality it loses to base. **The place to attack is the single point of the NSFW boundary.** General scenes / physical coherence and anime are weak on both — those are a different fight (LoRAs, etc.).

## What the NSFW A/B showed

- **Sulphur can render nudity / intimacy explicitly.** That's exactly the residual space the frontier won't enter. But **arm/anatomy breakage means seed gacha + prompt tuning are prerequisites.**
- Interestingly, **human action (e.g. boxing) holds up surprisingly well on Sulphur** — having trained on nude motion, its grasp of human movement seems *preserved* rather than degraded.
- **Two-or-more-person interaction currently collapses** — limbs mix between people. Intimate-couple scenes aren't usable yet. **Establish solo first, then expand to interaction.**
- **Vague prompts are also a failure mode.** Concrete descriptions (pose / framing / lighting / "photorealistic detailed skin") are required.

## The solo recipe that works

For a single subject, quality reaches a usable level. The established recipe:

1. **Restrict to a single subject**: `single ... alone in frame` in the prompt, `two people, multiple people, second person, crowd` in the negative.
2. **Concrete prompt**: spell out pose / framing / lighting / `photorealistic detailed skin`. Vague = collapse.
3. **Anatomy negative**: `deformed hands, mutated hands, extra fingers, extra arms, extra limbs, fused limbs, distorted anatomy, bad proportions, malformed body, ...`
4. **Higher resolution = higher quality**: 768×512 < 1024×576 < **1024×768** and up. Consistent with Sulphur being trained at higher resolution ("higher resolution is justice").
5. **Seed gacha still applies**: hunt a clean seed, fix the above, iterate.

## I2V is the mainstay — it structurally dodges the anatomy gacha

The decisive move against T2V's anatomy gacha is **I2V**.

**First build a still with perfect anatomy via text-to-image** (a solo full-nude where hands and proportions don't break). **Then animate it with Sulphur / 10Eros I2V.** Checking frame by frame, both **hold the source still's clean anatomy across several seconds (97 frames).** Because the starting point is already correct, T2V's gacha and breakdown are dodged by **path design.**

This is what overturned my earlier pessimism that "this isn't practical on 96 GB today." Not gacha-T2V, but **high-quality still → stable I2V** — that makes a v1 feasible on current hardware.

### The image_strength tradeoff (under exploration)

In I2V, the conditioning strength `image_strength` governs motion:

- **0.9 (strong conditioning)**: faithful to the source = stable anatomy, but **motion is modest.**
- **lower**: more motion, but higher **drift / collapse risk.**

"Low-motion solo I2V is usable with anatomy preserved" is confirmed. **Still untested: instruction-following (how well motion prompts are obeyed) and the limit of complex motion.** Sweeping image_strength down to add motion is the next step.

## Examples: 3-model comparison (base / Sulphur / 10Eros)

Same conditions, solo I2V across three models (768×1024 / ~4 s). **The stills are frames; play the clips via the links** (motion is the whole point).

### base (control: collapsed)

![base model I2V frame (collapsed)](/articles/i2v-uncensored-truth/base.jpg)

▶ [play base.mp4](/articles/i2v-uncensored-truth/base.mp4)

**It collapsed completely on this NSFW prompt.** base has strong general physical coherence, but it isn't **tuned** for this kind of content, so the I2V breaks anatomy. This is itself evidence for the thesis: **animating NSFW doesn't work on a general base model — you need an uncensored fine-tune.** Putting base here as a control makes it clear *why* the two below are necessary.

### Sulphur

![Sulphur model I2V frame](/articles/i2v-uncensored-truth/sulphur.jpg)

▶ [play sulphur.mp4](/articles/i2v-uncensored-truth/sulphur.mp4)

Where base collapsed, **this holds together.** An uncensored fine-tune that renders the NSFW boundary cleanly. Since this is a low-motion clip, physical breakdown isn't noticeable.

### 10Eros

![10Eros model I2V frame](/articles/i2v-uncensored-truth/eros.jpg)

▶ [play eros.mp4](/articles/i2v-uncensored-truth/eros.mp4)

An I2V-focused merge of the Sulphur line. **At low motion like this it's about on par with Sulphur**, also without breakdown. The gap between the two only shows under stronger motion (e.g. lowering image_strength) — still untested.

> Sulphur / 10Eros take the "clean still → I2V" path and hold the source anatomy across several seconds (the strength of not routing through T2V's gacha). base, from the same start, collapses — the watershed is **whether the model is tuned for NSFW.** Motion is on the low side here (image_strength ~0.9).

## What doesn't work yet (honestly)

- **Multi-person interaction collapses.** Solo-only is the realistic answer.
- **Complex motion / instruction-following is untested.** Confirmed only at low motion (image_strength 0.9).
- **image_strength sweep is pending.** Finding the motion-vs-stability sweet spot is the next stage.
- Hit rate is seed-dependent — a "gacha" element remains.

That's the honest current state. **Solo + low-motion I2V reached practical quality** — the v1 win. From here it expands toward motion and interaction.

---

*← [Back to the SFW version (strategy + experiment log)](/en/articles/residual-video-domain-i2v/)* ／ *This platform ([kotonia.ai](https://kotonia.ai)) is a solo-built, solo-run immersive voice × video roleplay service.*

(NSFW) An uncensored I2V experiment log: base / Sulphur / 10Eros comparison and the solo recipe.