12:15
AI video generation
The video argues that the core failure behind AI characters "falling apart" across scenes is that text-only descriptions force the model to guess appearance from scratch every time, and the fix is to give the model a fixed visual anchor — a three-panel character sheet saved as a reusable, taggable element (or, for real people, a Soul ID trained on many photos) — instead of retyping a description.
A single, very detailed character description (down to a scar's exact placement) still produces a visibly different-looking person across three different image scenes (rooftop bar, subway, night market).
The same description-only approach breaks down worse in video (Seedance 2.0) because the model generates dozens of frames, so the character can drift even inside a single shot.
Six total generations from one carefully written description came out with different hair, jaw, skin tone, and scar placement in nearly every one.
The claimed fix: build a 'character sheet' — one image with three panels (full body front, full body back, face close-up) on a plain gray studio background — so the model sees the character from every angle instead of inventing.
The plain gray background is presented as functionally important, not cosmetic: less visual noise around the character makes it easier for the model to keep him consistent later.
Save the character sheet as an 'element,' set its category to 'character,' and give it a name (e.g., 'Mateo Dino'); afterward, tagging @MateoDino in a prompt pulls in that exact face instead of writing a description.
Re-running the same three scenes with the tagged element produces a consistent character in both images and video, including the previously lost scar detail.
For multiple characters, build a separate character sheet and save a separate element for each one (demonstrated with Naomi Quinn and Marcus Nell).
Two or three tagged character elements can be combined in a single Seedance prompt to render multiple consistent characters together in one scene, including correct relative size differences between characters.
For scenes with back-and-forth dialogue, duration is extended (to 15 seconds) and the prompt is timed out second by second so dialogue happens in the right order.
Soul ID is described as a separate, unique Higgsfield feature for locking in the exact face of a real person: instead of one reference image, it trains on a whole set of 5–20+ photos of one person for the highest identity accuracy on the platform.
Soul ID training photos should be clear, from different angles, without duplicates, group photos, filters, sunglasses, or masks, since those confuse the model.
Soul ID currently only generates images, not video, and pairs with the dedicated 'Higgsfield Soul 2.0' image model.
Soul ID is framed as locking only the face/identity while letting outfit, setting, and lighting vary freely across generations.
The video's summary distinction: elements are the simpler, more flexible choice for made-up characters usable across both images and video, while Soul ID is for a real person who needs identical results across a lot of images.
GPT Image 2 — An image generation model in Higgsfield that the video calls 'the best model for a photo-real face that actually holds fine details.'. Apply: Select it in the image workspace with quality set to high, resolution to 4K, and aspect ratio 16:9 to generate both single-description character images and multi-panel character sheets.
Seedance 2.0 — A video generation model (referred to in the transcript as 'C Dance 2.0'/'seed ends'/'SeaArt 2') that the video says 'gives you the best results right now when it comes to consistency' and credits with 'preserving his facial identity with all the details.'. Apply: Use it for video generation (8–15 seconds, 1080p, 16:9), tagging a saved character element (e.g., @MateoDino) in the prompt instead of a written description to keep faces consistent across single- or multi-character shots.
Character Sheet — A single reference image with three panels — full-body front, full-body back, and a face close-up — all on a plain gray studio background, meant to show the model the character 'from every angle at once' instead of leaving it to guess from a description. Apply: Prompt GPT Image 2 to lay out these three panels on a plain background for each character before generating any scenes, then use the resulting sheet as the source image for a saved element.
Element (character element) — A saved, named asset (e.g., 'Mateo Dino') created from a character sheet and categorized as 'character,' which Higgsfield can pull into any prompt by tag instead of a written description. Apply: After generating a character sheet, save it as an element with category 'character' and a name, then tag @Name in image or video prompts to reuse that exact face/body, creating one element per additional character for multi-character scenes.
Soul ID — A Higgsfield feature 'built for locking in the exact face of a real person' that trains on 5–20+ reference photos of one person (instead of a single reference image) for the platform's highest identity accuracy, and currently generates images only, not video. Apply: Upload 20+ clear, varied-angle photos of one person with no duplicates, group shots, filters, or face coverings, name and train the Soul ID (about 10 minutes), then tag it in Higgsfield Soul 2.0 image prompts to reproduce that exact face across different settings, outfits, and lighting.
Higgsfield Soul 2.0 — The platform's dedicated image model for Soul ID characters, which the video says produces images that 'look a lot more realistic than with plain image generation.'. Apply: Switch the image generation model to Higgsfield Soul 2.0 whenever a prompt tags a trained Soul ID, to get the platform's most realistic, identity-locked results.
@ tagging — A prompting convention in which typing '@' plus a saved element's or Soul ID's name inserts that exact character into a generation in place of a written description. Apply: Replace character-description text in a prompt with '@CharacterName' (e.g., '@MateoDino,' '@NaomiQuinn') and describe only the scene/environment, combining multiple tags in one prompt to place several consistent characters in the same shot.
Second-by-second timed prompting — A prompting technique for multi-character video scenes where the prompt is timed out second by second so back-and-forth dialogue occurs in the correct order. Apply: For longer clips (the video uses 15 seconds) with two or more tagged characters exchanging dialogue, structure the prompt as a per-second timeline of what happens or who speaks, rather than a single freeform description.
The video uses the scar's disappearance as its proof point that description-only prompting is unreliable in principle, not just under-specified — even a detail the user explicitly spelled out couldn't survive across scenes.
Image consistency is framed as an easier, almost misleading benchmark; the video explicitly treats video generation as the real test because a character can shift within a single generated shot, not just between shots.
The claimed mechanism for why the character sheet works is about reducing what the model has to invent: showing it the character from every angle removes the guesswork that a description leaves behind.
Soul ID's design choice to separate 'identity' from everything else (outfit, lighting, setting) is presented as the source of its value — it explains why the same trained face works in a home office, on a beach, and in a gym without extra prompting effort for consistency.
The workflow implicitly reframes prompting itself for repeat characters: after the sheet/element is made, the user no longer writes appearance details at all, only environment and action, with the tag standing in for the whole description.
«Have you ever noticed that the biggest problem in AI creation is keeping your character looking like the same person in every scene?»
— 00:00
«the most important thing I've learned is that nothing makes a video look more fake than characters completely falling apart»
— 00:09
«If the model can't hold a scar that you literally spelled out for it, no written description will ever keep the same character across a whole video.»
— 03:12
«the way you actually fix this is to give the model one anchor to hold on to, instead of leaving it to guess from the description»
— 03:21
«these videos look like someone filmed the same person in different settings. They don't look like AI-generated videos of different characters»
— 05:58
«Soul ID trains on a whole set of photos of one person and that's what gives it the highest identity accuracy on the platform»
— 09:49
«An element is the simpler, more flexible pick for a character you build from scratch and use across both images and video, while Soul ID is the one you reach for when it's a real person you need identical across a lot of images»
— 11:38
Reception
Overwhelmingly positive reception with users expressing gratitude for practical tips, successfully applying them, and the creator actively engaging with supportive responses.
The video builds its case through direct before/after comparisons (same prompt, same scenes, with and without the character sheet/element) rather than assertion alone, which makes its consistency claims easy to follow, but every test is self-run by one creator inside a single platform (Higgsfield) with its specific models, so the results aren't independently verified or benchmarked against other tools.
12:15