Inline Studio released v1.2.71 with a Characters feature for FLUX.2: from one or two reference photos, a portable .char file with an identity fingerprint is created, allowing the same character to be generated in new scenes without training LoRA, adapters, or img2img. Consistency becomes measurable: after each frame, the system calculates a continuity score from 0 to 100.
What Happened
The new version of Inline Studio features a Characters panel for working with FLUX.2. From one or two reference photos, the system creates a .char file of a few MB in size; it contains the original references, a fixed text description, and the character's identity fingerprint, and the file can be copied between projects. Technically, references go through FLUX.2's native multi-reference channel: they are VAE-encoded into 32-channel latents, patchified (1024 tokens per 1 megapixel), and added to the noise-token sequence the model attends to at each denoising step. After each frame, a continuity score from 0 to 100 is calculated on the CPU using the formula 0.8 * face + 0.2 * subject: YuNet (approx. 230 KB, MIT) detects the face, SFace (approx. 37 MB, Apache-2.0) provides a 128-dimensional face embedding, and DINOv2-base (approx. 330 MB) provides a 768-dimensional object centroid; face similarity is compared with the nearest reference, not the average. The demo was rendered on FLUX.2 Klein 4B Base.
Context
Standard ways to fix a character in generative pipelines are training a LoRA on a specific person, adapters like PuLID or FaceID, or anchoring the face via img2img. Each of these paths requires either separate training or additional model passes. FLUX.2 can accept reference images through a native multi-reference channel, and the v1.2.71 mechanism is built exactly on this: it is in-context conditioning, where identity is transferred directly from reference images, not from a trained adapter. The new .char file can be considered a portable character card in the spirit of textual inversion — with the difference that the reference photos themselves are attached to the fixed description, and all identity transfer remains on the side of the base model.
Why This Matters for the Industry
The meaning of the release for the industry is in the product packaging, not the model itself: in engineering practice, this is the first ready-made training-free way to fix a character on FLUX.2. A team with a local GPU can now render single characters from one or two references and integrate the continuity score as a standard QC metric in the pipeline. Sellers of character LoRA training immediately face price pressure: the training-free alternative is free and ready instantly. If the format is polished and comparisons with baselines appear, the .char card could become a de facto standard for exchanging identity assets between projects in AI film pipelines on FLUX.2.
Why This Matters for Users
For those working with FLUX.2 or building AI video, the feature is available now: free, open source, locally on your own GPU. The workflow is simple: drag one or two photos into the Characters panel, select the character in the dropdown on the FLUX.2 node, and get the same person in new scenes — without reconnecting references for each frame and without describing the face in the prompt. Two practical limitations: there must be one person in the frame, and rendering cost increases significantly with the number of references — at 1024 px, five references increase the number of tokens by approximately 6 times at each denoising step.
What Is Still Unknown / Limitations
The demo was rendered only on FLUX.2 Klein 4B Base: the official guide's title also mentions a dev variant, but there is no quantitative data on it. There is no public quantitative comparison of the method with LoRA, PuLID, and img2img baselines, nor data on latency. Multi-person scenes are declared unsolved: the character's identity flows to other people in the frame. The final quality directly depends on how well the base model is trained to use the multi-reference channel. Furthermore, from the perspective of scientific novelty, the release is an engineering packaging of an existing FLUX.2 capability, not an advancement of the model itself.
Sources
Author
Look at AI, editorial team