> For the complete documentation index, see [llms.txt](https://docs.pletor.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.pletor.ai/how-to.../techniques.md). # Techniques #### Photography
How to achieve product consistency across generations * Give it several views, not one. A single reference forces the model to imagine every angle it wasn't shown, and it will happily hallucinate a back or a side that doesn't exist. You can Lock the hard facts in a product reference sheet. Views tell the model what the product *looks like*; * Write anti-drift guardrails. For each product, name what the model tends to get wrong and correct it explicitly in the prompt * To do so, you can either : * Create various views of your product to feed the product shot agent * See : T01 - Creating Product Views * And / or collapse views into one reference sheet (easier to maintain) * See : T02 -
How to upscale an image in 4K * You can use the upscale-image node, there are three model providers : * Topaz: the all-rounder. Use it to enlarge an image while staying faithful to the original, with no creative reinterpretation. It's the safe default for clean upscaling; 4x is the reliable sweet spot where detail still holds together. * Magnific: various upscalers. Creative upscaler will improve colors, Rather than just adding pixels, it reconstructs and "reimagines" new detail, textures, and richer color, steered by a text prompt and sliders (Creativity vs. Resemblance). * Enhancor: use it to enhance skin realism. Use it to fix the plastic, waxy look of AI-generated faces: it rebuilds realistic pores, fine hair, and natural skin texture. Reach for it on portraits and close-ups where skin realism is the priority * Use NBP or NB2 in 4K mode or Use Seedream 4.5, or 5 Lite in 4K mode * Using these models with precise upscale instruction can do the job right * With any of these, a well-written, precise upscale instruction does the job, describe what to preserve and what to sharpen, and let the model render the high-res version.
How to maintain consistent characters / models across generations * Character consistency is achieved by giving the downstream model various views of the same person. Including front, side views. Close-ups on the face, etc. If the character has specific teeth, make sure to show him smiling in the views. In general, try to give to models as much context as possible about the person. * To do so, you can: * Create a reference sheet of the face of your character * Create a reference sheet of the body of your model * Then give the model several images of the same character, various views and expressions, with a focus on their face
How to keep color, lighting & camera consistent across a set * Using one key frame to define color, lighting and camera quality accross all frames * Generate one "environment frame" that nails the target look: color grading, lighting setup, camera quality. Then pass it as a reference for every other frame in the set, but scope it explicitly in the prompt to *light, camera and color grading only* (locked instructions), so it doesn't bleed into composition or subject. * Pre-prod with the right prompts and same model * Same model for the whole set, and a shared LOCKED block across all prompts: identical camera body/lens, focal length, aperture, lighting description, grading keywords. Only the VARIABLE part (subject, action, framing) changes between frames. * Exemple : * Post-prod all the images with the same image reference. Using reference frames for color grading * Run all outputs through the same post-prod pass using a single reference frame for color grading. This catches the residual drift that steps 1–2 can't fully prevent. * See T12 - * See T11:
How to build and reuse credible scenes / environments * Generate one hero frame of the place with no subject (or minimal subject): this defines the architecture, materials, light direction and mood. This becomes the canonical reference for the location. * From the master frame, generate complementary views, reverse angle, side view, detail shots, using it as reference. This coverage set proves (and enforces) spatial consistency: recurring elements like windows, furniture or landmarks must reappear coherently across views. * Wide 3/4 angles show floor, walls and ceiling lines converging. This is what sells the space as a real, navigable volume rather than a flat backdrop. Reserve frontal or tight framings for inserts once the geography is established.
#### Video
How to build strong video agents * Storyboard first: generate each key frame as a deliberate shot (framing, angle, moment) rather than hoping the model invents good coverage. Every frame should answer "what does this shot do in the edit?" - establish, action, reaction, insert. Example here: * Chain clips into seamless continuity: use the last frame of a clip as the first frame of the next (or first/last frame conditioning where the model supports it) so motion, light and geography carry over without a visible cut. * When the sequence needs rhythm (chase, montage, product reveal), use models that generate multi-shot sequences in one pass: Seedance R2V or Kling 3.0 handle the cuts internally, which keeps energy and coherence better than stitching independent generations. * Finish with a 4K upscale pass on the assembled video. Generation resolution is a working format — upscaling recovers sharpness, texture and grain uniformity across shots, and masks minor inconsistencies between clips.
How to keep a face/character consistent across video shots * Assemble the character's canonical views: clean face shots (frontal + 3/4), full body, and outfit details. This sheet is the single source of truth, every generation references it, never a previous output (copies of copies drift). * Feed the reference sheet into Seedance's reference-to-video: strong identity retention across shots and camera angles. * Same principle via Kling's Elements: pass the character as a persistent element that survives across generations.
How to generate realistic UGC videos * Start with the script: hook, message, CTA, in the target language. It defines duration, pacing and tone, everything downstream (voice route, performance, even framing) depends on it. Keep it spoken-word natural, not ad copy: UGC that reads like a brief sounds like a brief. * Create a character - Define your creator persona once: face, age, styling, vibe. This is the identity anchor that everything downstream references. * Generate a realistic UGC frame of your character - Place the character in a credible UGC context using the UGC preset: phone-camera look, imperfect framing, natural lighting, real-world setting. This frame sets the amateur aesthetic that sells authenticity. T04 - (preset = UGC) * For the voice : * Native model audio: let the video model generate speech directly — Seedance 2.0 handles many languages well, and native audio comes with lipsync and performance for free. * ElevenLabs voice changer: generate or record a guide track, then transform it with an ElevenLabs voice for full control over timbre and identity — useful when the voice must match a defined persona or stay consistent across a batch. * Use voice changer for audio - T08 - * Generate a realistic voice and micro expressions Sync the character to the VO with lipsync and micro-expressions (blinks, head tilts, hesitations). This is where realism is won or lost — flat faces read as AI instantly.