Krea 2 Anygles: The Free AI Camera-Angle LoRA That Orbits a Person 360° From One Image

Every AI image creator hits the same wall: you finally get a character you love, then you need the same person from the side, from behind, or from a low hero angle, and the prompt “side view” gives you a stranger. Krea 2 Anygles, a free community LoRA that landed on Hugging Face on September 22, is the most direct fix I’ve tested for that problem. You give it one image of a person, drag three sliders (orbit, elevation, distance), and it re-renders the same figure, outfit, and scene from the new camera angle.

The free path is real: there’s a free browser demo on Hugging Face that needs no GPU, plus open weights and ComfyUI nodes for local use. The catches are hardware (it’s a Krea 2 workflow, so plan on roughly 21 GB of downloads), a revenue-capped license, and the fact that it only handles one human subject. Here’s what it does, how to run it today, and when the Apache-2.0 Qwen alternative is the smarter pick.

At a Glance

• What it is: a 229 MB rank-32 Control-LoRA for Krea 2 Turbo that changes the camera viewpoint of a person in an existing image.
• What’s free: the weights, the ComfyUI nodes, and a ZeroGPU web demo.
• What it costs: $0 to try; locally you need a GPU that can hold a 13.5 GB int8 Krea 2 model plus a 5.2 GB text encoder.
• Controls: full 360° orbit, elevation (±45° in the demo, ±60° in the tools), and distance from 0.6× to 1.8×.
• The one catch: one clear human only, and commercial use falls under the Krea 2 Community License (free under $1M annual company revenue).

What Krea 2 Anygles Actually Does

Most “multiple angles” tools are edit models that read a text instruction like “rotate 90 degrees” and guess. Anygles adds geometry. It first recovers a 3D body mesh from your photo using Meta’s SAM 3D Body, rotates that mesh to the camera position you asked for, and renders it as a surface-normal map, the same kind of map we covered in our Marigold V2 depth and normals post. That normal map is injected into Krea 2 as a spatial control, while your original image goes in as the identity and style reference.

The practical result: the body pose follows real 3D geometry instead of a text guess, and because every view starts from the original image (not the previous output), you don’t get the drift that builds up when you chain edits. According to the model card, it works across photography, painting, stylized 3D, and anime.

Specs, File Sizes and Hardware

Here is everything the ComfyUI workflow loads, with actual file sizes from the Hugging Face repos:

• krea2_anygles_rank32.safetensors: 229 MB
• krea2_turbo_int8_convrot.safetensors: 13.5 GB (the fp8 build is 13.1 GB; an nvfp4 build is 7.7 GB)
• qwen3vl_4b_fp8_scaled.safetensors: 5.2 GB
• qwen_image_vae.safetensors: 254 MB
• SAM 3D Body (dinov3): 2.1 GB, gated behind Meta’s SAM License

That’s about 21 GB total. The author doesn’t publish a VRAM number, so treat this as my estimate: with a 13.5 GB diffusion model, a 16 GB card is the practical floor, and 24 GB is comfortable. The workflow runs SAM 3D Body and Krea 2 one after the other, so they don’t need to share VRAM. The nvfp4 build could lower the bar on RTX 50-series cards, but the author only documents the int8 file, so that’s untested territory. Recommended settings: 8 steps, Euler sampler, simple scheduler, CFG 1.0, LoRA strength 1.0.

Krea 2 Anygles vs. the Free Alternatives

Anygles isn’t the only free camera-angle tool. The most popular one is fal’s Multiple-Angles LoRA for Qwen-Image-Edit-2511, and there’s a newer Blender-guided option for Qwen-Image-2.1 (covered in our Qwen-Image-2.1 breakdown).

The short version: Anygles is the most precise option for people, fal’s LoRA is the most flexible and the most permissively licensed, and AnyAngle gives you total control if you’re willing to open Blender.

A note on that comparison: the author ran it with fixed seeds and first attempts only, which is more honest than most showcases. Still, Qwen didn’t get Anygles’ normal-map guidance, and the author calls it “a practical feature comparison rather than an identical-conditioning ablation.” In this frame, Anygles clearly reaches the back-three-quarter view while Qwen stops near a side profile. No independent benchmark exists yet, so treat it as a promising demo, not a measured win.

Try Krea 2 Anygles Free

Option 1: no GPU, five minutes. Open the Krea 2 Anygles Space. Drop one full-body or half-body image of a single person into “Reference image,” set the Horizontal angle, Elevation, and Camera distance sliders (or click the Left −90° / Back 180° / Right +90° presets), and hit Generate one view. In my test, a cold Space took about 92 seconds to load its models the first time, then generated each view in about 19 seconds. It runs on Hugging Face’s free ZeroGPU tier, which gives each account a limited daily GPU allowance, so log in for more runs.

Option 2: ComfyUI on your own GPU. You need a current ComfyUI build with native Krea 2 support. Then:

cd ComfyUI/custom_nodes
git clone https://github.com/alexw5702-afk/krea2-anygles
cd krea2-anygles
pip install -r requirements.txt

# after accepting the SAM License on Hugging Face
hf download facebook/sam-3d-body-dinov3 --local-dir ComfyUI/models/sam3d_body

Put the Anygles LoRA in models/loras, the Krea 2 Turbo int8 file in models/diffusion_models, the Qwen3-VL 4B encoder in models/text_encoders, and the VAE in models/vae. Restart ComfyUI and drag the krea2_anygles_workflow.json onto the canvas. The four custom nodes are Krea2 Anygles Camera, Krea2 Anygles Encode, Krea2 Anygles Load LoRA, and Krea2 Anygles Model Patch. Use the included loader: the model card warns that ordinary LoRA loaders and the auto-generated Diffusers snippet silently drop the spatial-control tensor.

Option 3: Python. The repo ships prepare_normal.py (builds the target normal map) and example.py (generates the view). You’ll need to accept the access terms for Krea-2-Turbo and run hf auth login first.

Honest Trade-Offs

• People only. The release supports one clear human subject. Animals, products, crowds, and full-scene reconstruction are out of scope. For those, use fal’s Qwen LoRA or a 3D route like Mira-Scene’s image-to-3D scene model.
• Hidden sides are invented. The back of a jacket, logos, text, fingers, and accessories can change between views. It’s a viewpoint generator, not a scanner.
• Heavy setup. About 21 GB of downloads and two gated models (Krea 2 Turbo, SAM 3D Body), each with its own license click-through.
• License stack. The Anygles weights inherit the Krea 2 Community License: you own your outputs, and commercial use is allowed only if your company makes under $1M a year. SAM 3D Body carries Meta’s SAM License, and the pipeline code is Apache 2.0.
• Unofficial. This is a single developer’s release with around 770 downloads so far, not endorsed by Krea or Meta. Expect rough edges and slower fixes.

Workflow Tips

• Always regenerate from the original. For a turnaround sheet, render each angle from the same source image. Never feed a generated view back in.
• Give SAM a clean body. Full or three-quarter body shots with visible limbs recover far better meshes than tight headshots or heavily cropped poses.
• Make your source with Krea 2 itself. A matching base model keeps style drift low. If your Krea 2 Turbo outputs all look alike, our Krea 2 Turbo diversity LoRA fix helps before you start rotating.
• Use the optional prompt for lighting only. Things like “soft afternoon light” get appended after the camera instruction. Don’t restate the pose or the angle.
• Push angles in small steps first. Large elevation and distance changes can shift proportions and background layout. Test at ±20° before trying ±45°.

Want to see how the Qwen side of this comparison performs on its own? Watch our Qwen Image 2.1 review, and subscribe to the Creative AI Show YouTube channel for hands-on tests of new image tools.

FAQ

Is Krea 2 Anygles free?

Yes. The LoRA, pipeline code, and ComfyUI nodes are free to download, and the Hugging Face Space lets you generate views in a browser at no cost within ZeroGPU’s daily limits.

Can I use Krea 2 Anygles commercially?

Yes, if your company’s total annual revenue is under $1 million, under the Krea 2 Community License. Above that, you need a Krea enterprise license. If that’s a problem, fal’s Qwen-Image-Edit-2511 angles LoRA is Apache 2.0.

What are Krea 2 Anygles’ system requirements?

There’s no official VRAM figure. The int8 Krea 2 Turbo model is 13.5 GB, so plan on a 16 GB NVIDIA GPU at minimum (24 GB is comfortable), about 21 GB of disk space, and a current ComfyUI build with native Krea 2 support.

Does Krea 2 Anygles work on objects or animals?

No. It depends on a human body mesh, so it’s built for one clear person. For products, pets, or scenes, use the Qwen-Image-Edit-2511 Multiple-Angles LoRA instead.

Accurate as of September 29, 2026. Model versions, licenses, and free-tier limits change quickly.