Multi-angle AI can move the camera. Not every way yet.
Multiple-Angles and Next Scene LoRAs turn one frame into new camera angles and the next shot of a sequence. We ran the camera prompts on game art: the close-up and the high angle held, the rotation didn't.
Two community add-ons for Alibaba's Qwen-Image-Edit model now move the camera inside a picture you already have. Multiple-Angles re-renders the same subject from a new position: closer, higher, wider or rotated. Next Scene draws the next shot of a sequence from one frame, as if the camera had kept rolling. For game ads, that's the step between one good piece of key art and a video built from several shots of the same character. We ran the camera prompts on our own game art. The close-up and the high angle held up. The rotation didn't move at all.
What the Multiple-Angles and Next Scene LoRAs do#
Both are LoRAs for Qwen-Image-Edit-2509, the open image-edit model Alibaba's Qwen team released in September 2025 under Apache 2.0. A LoRA doesn't replace the model. It's a small set of extra weights that teaches the model one specific behaviour, and anyone can train one and publish it.
The Multiple-Angles LoRA by the developer dx8152 teaches camera moves. Its model card lists 9 example instructions, written in Chinese: move the camera forward, left, right or down, rotate it 45° left or right, switch to a top-down view, a wide-angle lens or a close-up. "There are no trigger words," the card says, and it invites you to try other moves. It was downloaded 81,936 times in the month to 18 October 2026.
The Next Scene LoRA by lovis93 does the other job. Its model card describes "cinematic frames with natural visual progression": you give it a frame, start the prompt with "Next Scene:", lead with a camera direction, and it draws the following shot. Dolly in, pull back, track the subject, reveal the environment. The creator is clear about where it doesn't fit: static portraits and non-sequential edits. It's meant for scene-to-scene transitions rather than detailed object manipulation.
| Multiple-Angles (dx8152) | Next Scene (lovis93) | |
|---|---|---|
| What it changes | Where the camera stands | What happens in the next shot |
| What stays the same | The moment: same pose, same scene | The world: same character, place and light |
| How you ask | A camera instruction, e.g. rotate 45° left | "Next Scene:" plus a camera direction |
| Base model | Qwen-Image-Edit-2509 | Qwen-Image-Edit-2509 |
| Licence | Apache 2.0 | MIT, with attribution asked for commercial use |
| Use in an ad | Another angle of the same shot | The next shot in a storyboard |
From nine camera prompts to a 96-pose grid#
A list of nine phrases is a start, not a camera rig. The community's next step was to turn camera position into numbers.
fal's Qwen-Image-Edit-2511 Multiple-Angles LoRA covers 96 poses: 8 horizontal directions in 45° steps, 4 heights from a −30° low angle to a 60° high angle, and 3 distances from close-up to wide. It was trained on more than 3,000 renders of 3D scenes. fal's hosted version takes the angles as plain numbers (horizontal 0–360°, vertical −30° to 90°, zoom 0–10) at $0.035 per megapixel, according to its API docs.
| Multiple-Angles (2509) | 9 |
|---|---|
| Multi-Angles (FLUX.2) | 72 |
| Multiple-Angles (2511, fal) | 96 |
A FLUX.2 version by lovis93 lists 72 poses with the same kind of trigger-word prompt. Check the licence before you build on it: the FLUX.2 [dev] base model carries the FLUX Non-Commercial License.
Meanwhile the base model moved too. When Qwen released Qwen-Image-Edit-2511 in December 2025, its notes listed less image drift, better character consistency and "integrated LoRA capabilities". On viewpoints they say "generating new viewpoints can now be done directly with the base model". Qwen doesn't say which community LoRA, if any, it learned from, so we won't either.
We ran the camera prompts on game art#
We wanted to know what the base model does with a game character today, without a LoRA. We took sample art we generated for our own site in a Clash of Clans style (not a customer's art and not Supercell's): a barbarian on a village background. We sent it to the base Qwen-Image-Edit-2511 model through our image provider, with four of the Multiple-Angles LoRA's camera instructions and one sample each. Each image cost about two cents. We didn't run the LoRAs themselves, because our provider doesn't host them.
Two prompts worked well. The close-up kept the face, the helmet and the moustache, and softened the village into a believable background. Asked for top-down, it gave a high angle: the camera above him, with the grass and buildings re-laid around that view and the character's pose intact.
Two prompts did nothing. "Wide-angle" returned a near-identical frame. "Rotate the camera 45° left" did too, in Chinese and again in English. That's in line with what users report on the Multiple-Angles LoRA's own discussion board: threads about left and right getting mixed up, and about having a bird's-eye and a worm's-eye view with nothing in between.
One more report matters for game teams. A user turning game prop icons to a front view said accuracy was "only 50%". The author replied that this is perspective correction, a different task that needs its own training. Icons, items and UI aren't scenes. Don't expect a scene-trained camera LoRA to fix them.
Why game ads need the camera to move#
A video ad is a short sequence of shots of the same character. The hook opens on one framing, the gameplay moment cuts to another, the reward lands on a close-up. If every shot is generated separately, the character drifts: the helmet changes shape, the colours wander, the village moves. That drift is one of the things that makes AI game ads look generated.
Image-to-video models take frames as inputs too. Google's Veo 3.1 docs, for one, describe a first frame, an optional last frame "for interpolation" that needs a first frame with it, and up to 3 reference images to keep a character or product consistent (reference images need an 8-second clip). A re-angled frame is a first frame for the next shot. A Next Scene frame is a storyboard panel. Both start from art the studio already approved.
- 01Approved key artthe studio's character and world
- 02New anglesclose-up, high angle, side
- 03Next shotsthe storyboard, frame by frame
- 04First frameschecked against the original art
- 05Video clipsimage-to-video per shot
- 06The cutedited, captioned, sized
Static ads benefit as well. A character facing the wrong way for a layout where the button sits on the right. A low angle that makes a boss look bigger in a 9:16 story. A close-up for a 1:1 feed card where the full-body pose reads too small. Today those mean a new render from the art team or a crop that loses the frame. Our guide to mobile game ad sizes explains why cropping a master breaks layouts. When re-framing works, it keeps the frame a crop would lose.
Where it breaks today#
The honest list, from the model cards, the discussion boards and our own test:
- Unseen sides. Rotations that reveal the side or back of a character are the weakest moves. The model has to invent what the art never showed.
- In-between angles. Fixed instruction lists jump between views. The 96-pose grid helps, but its card publishes no accuracy numbers.
- Version lock. A LoRA is trained for one base model. Users report the 2509 Multiple-Angles LoRA "destroys the image" on 2511. Every base-model upgrade can break a pipeline built on it.
- Invented detail. A close-up has to add detail the source didn't have, such as rivets, teeth or fabric. It looked right on our barbarian, but it's still the model's guess, and the art director has to sign it off.
- Licences. Qwen's models and the dx8152 and fal LoRAs are Apache 2.0. Next Scene is MIT and asks commercial users to test it and credit it. Anything built on FLUX.2 [dev] needs a commercial licence check.
None of this makes the tools useless. It makes them production steps that need a check, like any other.
How we'd build it into Vidscape#
This isn't in the product today. Here is how we'd want it to work, and the rules we'd hold it to.
Vidscape already builds video ads scene by scene from a game's own art: you approve the script and the first frame of each scene before any footage is generated. Camera edits would make those first frames different angles of the same world, rather than separate pictures that happen to look alike. The same move inside a static ad would make the angle a property of the character layer, so turning the barbarian toward the button would be an edit. That's the same principle behind our editable image generation: the model does the production, the team keeps control of the result.
- 1Start from approved art
The studio's own character and key art would be the reference, never a text description of them.
- 2Pick the camera with controls
Direction, height and distance would be settings, like the numeric angles fal's hosted version takes, so the same move can be repeated across a batch.
- 3Check identity against the original
The new angle would sit beside the source art before it goes anywhere, so a drifted helmet is caught before the ad review.
- 4Use it as a layer or a first frame
A re-angled character would drop into a static ad as a swappable layer, or start the next shot of a video.
If you're a creative director weighing how much of this to let into a pipeline, our page for creative directors covers how we think about AI and art direction. We'll write about camera edits again when we've run them on more than one character, including the LoRAs themselves.
Frequently asked questions#
What is the Multiple-Angles LoRA?
An add-on for the Qwen-Image-Edit image model that re-renders a picture from a different camera position: closer, wider, higher, lower or rotated. The original by dx8152 targets Qwen-Image-Edit-2509, and its example instructions are in Chinese. A newer version from fal targets 2511 and covers a grid of 96 camera poses.
What does the Next Scene LoRA do?
It draws the next shot of a sequence from one frame: a push-in, a pull-back, a tracking shot or a reveal of the environment. Prompts start with "Next Scene:" and lead with a camera direction. The creator says it suits storyboards and scene-to-scene transitions, and isn't optimised for static portraits or non-sequential edits.
Does Qwen-Image-Edit-2511 need a LoRA for new angles?
Qwen says new viewpoints work with the base model. In our test the base model handled a close-up and a high angle, and ignored a 45° rotation and a wide-angle instruction. A dedicated camera LoRA may still do better on rotations.
Can I use these models commercially?
Qwen-Image-Edit, the dx8152 LoRA and the fal LoRA are Apache 2.0. Next Scene is MIT, and its creator asks commercial users to test it and give attribution. FLUX.2 [dev], the base of another multi-angle LoRA, has a non-commercial licence. Check each one before you ship.
Does Vidscape re-angle characters today?
No. Vidscape builds video ads scene by scene from a game's art, with each scene's first frame approved before footage is generated. Camera edits are something we're exploring. They aren't a feature yet.
- Qwen-Edit-2509-Multiple-angles (model card) — dx8152, Hugging Face · Nov 2025
- Qwen-Edit-2509-Multiple-angles discussions — Hugging Face · 2025–2026
- next-scene-qwen-image-lora-2509 (model card) — lovis93, Hugging Face · Oct 2025
- Qwen-Image-Edit-2511-Multiple-Angles-LoRA (model card) — fal, Hugging Face · Retrieved Oct 2026
- Qwen Image Edit 2511 Multiple Angles API — fal · Retrieved Oct 2026
- Flux-2-Multi-Angles-LoRA-v2 (model card) — lovis93, Hugging Face · Dec 2025
- FLUX.2 [dev] (model card and licence) — Black Forest Labs · Nov 2025
- Qwen-Image-Edit-2509 (model card) — Qwen, Hugging Face · Sep 2025
- Qwen-Image-Edit-2511 (model card) — Qwen, Hugging Face · Dec 2025
- Qwen-Image (repository) — QwenLM, GitHub · 2025–2026
- Generate videos with Veo 3.1 in the Gemini API — Google AI for Developers · Retrieved Oct 2026
