Qwen Image 2512 LoRA Review: Load Your Own Weights, No Filter

The gallery holds published, safe-for-work examples from the Qwen Image 2512 LoRA model page. Our content-ladder test images aren't shown; we report the results only.

A photo LoRA that reads like a film still. A real-life photo LoRA on text-to-image, two women slow dancing barefoot in a dim apartment at blue hour, one in a champagne silk slip with a strap sliding off her shoulder, the other in an open white shirt. The cyan window light and the pink lamp land where the prompt put them, and skin, silk and cotton come out unpolished, the way a camera sees them rather than a beauty filter.

A different palette, same reliability. A watercolour-anime LoRA on the same base, two women on a balcony at sunset with wine glasses in hand, loose washes and paper texture taking over from the photographic look.

A pixel-art trigger word. "Pixel Art" at the front of the prompt turns the same base into a retro arcade scene at 16:9, without losing the two figures' poses or expressions.

A new angle from one photo. The editing endpoint's multiple-angles LoRA took a side view of a man on a hotel balcony at night and re-shot him as a front-facing close-up, with the same face, the same open black shirt and the same skyline behind him, from a single source image. Only the camera moved.

Illustration into photograph. A photo-realism edit LoRA at 0.85 turned a drawn singer at a microphone into a photograph, holding her curls, her green velvet dress and the stage light in place.

On text-to-image, put the LoRA's trigger phrase first if it has one, then describe the subject, setting and light in full sentences; Qwen-Image reads long, concrete prompts well. On editing, use the LoRA's own wording when it expects one, and otherwise open with what each numbered picture contributes before saying what should change.

1. A photo LoRA with the light written out (text-to-image, 3:4). The exact prompt behind the slow-dance example, real-life photo LoRA at strength 1. This LoRA has no trigger phrase, so the prompt opens with the kind of shot and then names both light sources and their colours.

cinematic film still of a lesbian couple, two beautiful women in their early thirties slow dancing barefoot in a dim apartment at blue hour, one in a champagne silk slip dress with a thin strap slipping off her shoulder, the other in an oversized white shirt worn open over a black bralette, foreheads touching, eyes closed, soft smiles, her hand on the small of the other's back, tall window with city lights out of focus behind them, warm pink lamp light on the wall, cool cyan window light on their skin, shallow depth of field, 35mm film grain.

Run this prompt on Qwen Image 2512 LoRA →

2. The LoRA's own fixed phrase (image editing, one source). The exact prompt behind the camera-angle example; the "<sks>" opener is what this particular LoRA was trained to read, copied from its model card. Keep it, and change only the view words after it.

<sks> front view eye-level shot close-up

Run this prompt on Qwen Image 2512 LoRA →

3. A plain instruction, no fixed phrase (image editing, one source). The exact prompt behind the illustration-to-photograph example, LoRA at 0.85.

Transform this illustration into a realistic photograph: the same singer with dark curled hair in a green velvet dress at a vintage microphone, stage lighting, natural skin.

Run this prompt on Qwen Image 2512 LoRA →

On 27 September we ran our content ladder against both endpoints with no LoRA loaded, three runs per probe on each: L1 Suggestive, L2 Mature themes, L3 Strong, L4 Intense and both L5 Extreme probe types (restraint and impact, and gore). On each endpoint fifteen of eighteen runs rendered as written. The exception was L4 Intense both times: the three text-to-image runs implied the act and kept one partner's clothing on, and the three edits posed the couple close together without showing the act. Neither endpoint refused or swapped the scene for something else.

On 26 September we loaded a third-party anatomy-focused LoRA on text-to-image and a different adult LoRA on editing, one run each. Both endpoints rendered the L4 Intense probe as written, editing starting from a clothed reference image. The page label is Uncensored: we add no content filter on our side, and neither the prompt, the LoRA weights nor the output is reviewed.

A refused request is not billed; none of our runs across either round was refused.

Not tested yet: the L4 Intense gap re-run with a LoRA at a lower strength; scenes with two women or two men; real photographs as text-to-image subjects; a LoRA-loaded run at every aspect ratio.

Adults only, in every prompt, source image, LoRA and output. Never depict a real, identifiable person without their consent.

Both endpoints take up to three LoRA entries per request, each a direct HTTPS link to a .safetensors file, with an independent strength per entry. Text-to-image LoRAs are trained for Qwen-Image-2512; editing LoRAs are trained for Qwen-Image-Edit-2511. Loading one on the wrong endpoint costs the same $0.03 and changes almost nothing, so the first run with any new file is a check, not a final: use the trigger phrase from its model card if it has one, render one image, and look for an obvious effect.

A real-life photo LoRA at strength 1, the women's version of a pair that also comes in a men's version, turned a written-out blue-hour scene into something that reads like a film still, with faces and hands still holding together in a close two-person pose.

A pixel-art style LoRA, triggered by the phrase "Pixel Art" at strength 1, rendered a different scene as a retro arcade frame at a wide 16:9.

A multiple-angles edit LoRA at strength 1, with its fixed "<sks>" phrasing, re-shot a single side-view photo from the front while keeping the face, shirt and skyline unchanged.

A photo-realism edit LoRA at 0.85 turned an illustrated singer into a photograph, holding her hair, dress and pose in place while the medium itself changed.

Every one of these ran with a LoRA hosted outside SpicyAPI: point path at any reachable .safetensors link trained for the matching base, including files people share on Hugging Face or Civitai for these two models. Weights load with no review of what they contain or what they were trained to draw, which is also what let an adult LoRA close our L4 Intense gap on both endpoints, as described above.

Qwen-Image-2512 launched to strong open-source reception: "Qwen's latest Qwen-Image-2512 is currently the strongest open source model." wrote one commenter on its release day, pointing to community GGUF conversions for running it locally. The community LoRA ecosystem built up fast too, though it comes with its own quirks: one widely used turbo-speed LoRA "causes Qwen 2512 to make everything realistic," and its own discussion thread notes "the only way to make anything animated is to use a step count of 2 or less." That matches our own advice above: check any new LoRA's effect on one image before trusting it in a batch.

• A LoRA trained for the older Qwen-Image line? Qwen Image 2.1 LoRA runs the same up-to-three-LoRA setup on Qwen's earlier open release, from $0.03 at 1K up to $0.102 at 2K. See our Qwen 2512 LoRA vs Qwen 2.1 LoRA comparison.
• Community LoRAs on a different base model? Z-Image Turbo LoRA and FLUX.1 Dev LoRA load the LoRA hubs built for those two models instead, at $0.012 and $0.018 a still.
• No LoRA to load at all, still uncensored? Qwen Image Spicy Edit is a preset uncensored editing endpoint at $0.038 an image, no weights file required.