Skip to content

Angle builder#

The angle builder generates the standard reference views of a character from the reference face: front, back, left profile and right profile. These angles are what a LoRA dataset is built from and what a video shoot can use when a shot needs a specific view.

The angle generation panel

  1. Open the character and click Create angle slots if the five cards are not there yet: face close-up, front, back, left and right.
  2. On a card, click Generate with AI. The generation panel opens with the face close-up (or, when there is none, the base reference) as its source. Pick the pipeline, the models it needs and optionally adjust steps, CFG, size, seed and the character description that is worked into the angle prompt. Generate queues the image; when it completes it is attached to the card. See Choosing a pipeline below.
  3. Repeat for the other angles. Pick from gallery assigns an existing image instead, for example a photo you already have of that view; photos uploaded to the Assets tab are listed next to generated images.

Generating all five angles in one go is not available yet; the Generate all button only announces it.

Choosing a pipeline#

  • FLUX.2 Klein (reference image) — recommended. The reference face is passed to a FLUX.2 Klein model as a reference image, so the generated views keep the same face without an IP-Adapter. Prompts are natural-language, single-person descriptions of each view; the character description is inserted after "this person". Defaults: 4 steps, CFG 1, 1024×1024 for the face close-up and 832×1216 for the body views; sizes are rounded to multiples of 16. Single-file Klein checkpoints need a standalone FLUX.2 VAE and a Qwen3 encoder record (Klein 9B pairs with Qwen3 8B, Klein 4B with Qwen3 4B); the panel picks them automatically when there is exactly one candidate, and Diffusers-format models carry both themselves. A Klein 9B model in bf16 needs about 18 GB, so on a 16 GB card it is offloaded and an image takes 30-90 seconds. The model stays loaded afterwards; use Clear model cache in the queue panel before a long video job to free VRAM.
  • SDXL + IP-Adapter — the reference goes through an IP-Adapter, which transfers style and rough likeness rather than identity; expect the face to drift. The IP-Adapter must match the image encoder installed with it: the plain SDXL adapter pairs with the IP Adapter SDXL Image Encoder, the ViT-H variants need the ViT-H encoder. A mismatch fails with "Selected CLIP Vision Model is incompatible with the current IP Adapter". Keep the description specific (clothing included); the default negative prompt already guards against collages and unwanted nudity.
  • SDXL + Z-Image Refiner — the SDXL result plus a Z-Image detail pass; same identity limits as SDXL.

Judging the result#

Compare each angle with the base reference at the same size. Reject an angle when the face shape, eye spacing or hairline changed, not when the light or expression differs. The back view can only be judged on hair and build. Regenerate rejected angles with a new seed; a small adjustment to the description helps when every seed shows the same drift.

Using angles elsewhere#

  • Training Studio builds the character dataset from these angles plus the base reference; see Datasets.
  • In Video Studio, pick an angle as the face reference when the shot needs that view.