Wan 2.2 Spicy LoRA Review: Your Own Weights, Uncensored

The gallery holds published, safe-for-work examples from the Wan 2.2 Spicy LoRA model page, shown as rendered. Our ladder test clips aren't shown; we report the results only.

Three LoRAs at once, no averaging. With a style LoRA, a fabric LoRA and a lighting LoRA all loaded, she turns on the spot and the satin catches a dark gold key light; each LoRA's contribution reads distinctly rather than blurring into a single look.

A continued take, same set carried over. The crypt clip continues an earlier shot with the identical LoRAs attached: candle flames sway, she turns her head and the sheer sleeve slides down her forearm, and the palette and grain match the clip it extends from exactly.

The ladder clips, as results only. With no LoRA loaded, our L1 through L3 stills and one L5 Extreme probe type came back at the level asked on image-to-video. L4 Intense softened in all three runs; one implied the act rather than showing it, two replaced the requested action outright. The L5 Extreme restraint/impact probe split evenly between rendering as asked and softening across two test dates. On video-extend, single runs at every rung, including L4 Intense and both L5 Extreme probe types, all came back exactly as asked.

The still (or the source clip, for video-extend) sets the content; the prompt names the beats and, when LoRAs are loaded, can call out what each one should contribute. Keep the wardrobe and setting details out of the prompt if the still or source clip already shows them.

1. Three LoRAs stacked, one beat (720p, 5 s). The exact prompt behind the stacked-LoRA clip. Naming what the LoRAs add ("the satin catches the dark gold key light") helped keep their contributions distinct rather than blended.

With all three LoRAs stacked - style, fabric and lighting - she turns on the spot and the satin catches the dark gold key light. The satin slides off one shoulder as she turns, the dark gold key raking the bare back. Cinematic, shallow depth of field, heavy film grain.

Run this prompt on Wan 2.2 Spicy LoRA →

2. Continuing a take with the same LoRAs (480p, 8 s, video-extend). The exact prompt behind the extend clip, written as a continuation rather than a fresh scene.

She turns away from the light and the satin falls off the other shoulder. Continue the clip for another eight seconds with the same LoRAs attached: the stillness breaks and she turns away from the light.

Run this prompt on Wan 2.2 Spicy LoRA video-extend →

3. A flat illustration brought to life (720p, 5 s). The exact prompt behind the pool clip, written to keep the source image's flat print look rather than push it toward photorealism.

The illustration comes alive while keeping its flat airbrushed print look: the pool water ripples and laps at her legs, the neon arcs pulse and their reflections wobble across the surface, water keeps running off her hair and shoulders. She lowers her chin from the sky, opens her eyes and turns her face toward the viewer, then draws one hand up out of the water and slowly pushes her wet hair back.

Run this prompt on Wan 2.2 Spicy LoRA →

More Wan prompts, by scene, are in our uncensored AI video prompt guide.

On 25 September 2026 we ran our content ladder against both tasks. On image-to-video, with no LoRA loaded, three tries per rung: L1 Suggestive, L2 Mature themes and L3 Strong came back exactly as asked, nine runs, nothing softened. L4 Intense softened all three times, twice replacing the requested action and once implying it instead of showing it. The L5 Extreme restraint/impact probe rendered on this date and one earlier date, splitting 2-for-4 between rendering and softening across the two dates; the L5 Extreme gore probe rendered all three times. On video-extend we ran one probe per rung (two for the restraint/impact probe, across two dates): every one of the seven runs came back exactly as asked, including L4 Intense and both L5 Extreme probe types, though that record is thinner than image-to-video's. The page label is Uncensored: we add no content filter on our side. The Wan line side by side is in our Wan uncensored test.

Freedom Score: on this page's score card, recalculated as runs are added.

Make the still (or source clip) with a model that goes as far as you need; this endpoint's LoRA fields change style and motion, not what's already in the frame.

Not tested yet: a request loading two or three LoRAs at once against the harder rungs (our ladder runs used no LoRA); 720p on the ladder (every ladder run we inspected was 480p); the 8-second duration on image-to-video's ladder runs; real photographs as the still; anime and illustrated stills at the harder rungs.

Adults only, in every still, LoRA, prompt and output. Never depict a real, identifiable person without their consent, and only load LoRA files you have the right to use.

• No LoRA to load? Wan 2.2 Spicy costs less per second, adds a closing frame, and passed our L1 through L4 runs outright. Our Spicy review.
• Only need image-to-video, not video-extend? Wan 2.2 LoRA charges the same $0.024 at 480p for the same LoRA fields on that one task.
• Base Wan 2.2, for comparison? Wan 2.2 image-to-video is a flat $0.24 for a fixed, silent 5 seconds and carries the Softened label, with no LoRA input at all.
• LoRAs on a different family? MiniMax H3 LoRA takes LoRAs on text-to-video, image-to-video and reference-to-video, at MiniMax's own content policy rather than Wan's.
• Need the hardest rung to render rather than soften, and don't need LoRAs? Wan 3.0 rendered our L5 Extreme probes on 15 September, from $0.045 per second at 480p.