the voice is the one thing that still gives ai video away and i just KILLED it. when your ai model's voice is made separately from the video, it sits on top of the scene. sounds like someone pasted it in the fix was writing the sound into the same prompt as the video. her line, the footsteps, the fence rattling when she backs him into it now her voice gets louder when she's inches from the lens and quieter a few feet back. the room stays quiet underneath, just open air and distant traffic i ran the audio through gemini and it read it as an authentic phone recording. every sound in it came back as real rewatch your last clip. if her voice sounds the same at every distance, write the sounds into the prompt instead of adding them after

you can run a whole ai influencer setup for next to nothing everything you need to get her onto fanvue runs on free models inside comfyui, and you're allowed to use all of these commercially - z-image turbo makes the photos. photoreal stills on a 16gb card, and the fp8 version fits on 8gb - qwen-image-edit keeps her face the same across photos, no training needed - wan 2.2 turns a photo into video. add the lightx2v lora and it takes about 6 steps instead of 20 - qwen3-tts clones her voice from 3 seconds of audio - seedvr2 upscales the photos and the video one 10-second clip on a paid video model runs $2.30 before failed takes. that's more than six hours of a rented 4090 so if you don't own a card, rent an rtx 4090 on runpod for $0.34 an hour, billed by the second, and load comfyui on it

i never sit down wondering what to post.. when one of my ai model's reels does well i write down three things. the scenario, the framing and the length one of mine is her in the driver seat, same angle every reel. the only thing that changes is the question those three lines go into claude and i ask for 100 topics in the same format my page posts 3 a day, so 100 topics is 33 days of content from one winner to find the winner, check each reel at the 3-hour mark and again at 24 hours. if it gets double your usual views, lock that format before you write the 100, test two versions of it that change one thing each, just the question or just the length take your best reel tonight, write its scenario, framing and length in three lines and ask claude for 100 topics in that format

screenshot any photo you like and this prompt writes you the exact instructions to recreate it: you write image prompts for one ai model. she's always the same girl, so her face never comes from the text. it comes from her reference image. how this works: i send you a photo of a scene i want her in (a pinterest pic, a screenshot from a reel, anything), or just a line like "her at the gym". you send back one json prompt that recreates that scene with my model in it. nothing else, no explanation. rules: 1. her face and body come from the reference image i attach in the generator. never describe her face, eye colour or bone structure. describe her body only in neutral words (curvy, slim, athletic) and let the reference do the rest. 2. copy the scene, not the person. take the location, outfit, pose, framing and light from my photo. ignore whoever is in it. 3. copy the colour grade. read the white balance, contrast, saturation and the light source off my photo and write them into the json, so she doesn't come out with the flat grey ai look. 4. it's always a phone photo. shot on an iphone, vertical 9:16, natural light or phone flash, candid, like she didn't know it was being taken. never a studio, never a dslr, never cinematic. 5. one imperfection, the same one every time: [her imperfection, e.g. a small mole under her left eye]. plus normal skin: visible pores, real texture, slight sensor grain, no retouching. 6. if my photo is vague or it's only a line, fill in what's missing yourself: where she is, the time of day, what she's wearing, who's holding the phone. 7. the seven ai tells and anything that looks like a pro camera go in the negative prompt every time. 8. 1080p, never 4k. reply with this json and nothing else: { "identity_lock": true, "reference": "her reference image: face and body", "scene": { "location": "", "time_of_day": "", "outfit": "", "pose": "", "framing": "mirror selfie / front camera selfie / a friend took it" }, "camera": { "device": "iphone, main camera", "aspect_ratio": "9:16", "resolution": "1080p" }, "light_and_grade": { "light_source": "", "white_balance": "", "contrast": "", "saturation": "" }, "realism": { "imperfection": "[her imperfection]", "skin": "visible pores, real texture, no retouching", "grain": "subtle phone sensor grain" }, "prompt": "everything above as one line, ready to paste", "negative_prompt": "plastic skin, empty eyes, waxy face, airbrushed skin, perfect teeth, warped hands, studio lighting, dslr, cinematic, bokeh" } wait for my first photo.

f*ck it.. here's the exact structure i break every viral reel into before i remake it: give gemini any clip with this prompt: "describe the video frame by frame, scene by scene, with an accurate description of the story and what happened" the breakdown only works if every timestamped block fills the same fields: 1. the shot. camera move and distance, "wide POV, eye level, handheld walk" 2. performance. body language and energy, "sits still" 3. dialogue with delivery. the stammer written out word for word 4. sound. the specific noise in that block, footsteps then the loud metal crash of the chain-link fence 5. her look. repeat it every block, black graphic top, ripped black tights, grey furry boots, silver pendant 6. find the turn. gemini names the exact second the story flips, 0:08 she stands and backs him into the fence gemini writes a video prompt at the end with [MY MODEL] as a placeholder push that prompt into seedance with your ai model as the reference, or a scene board of her you made in chatgpt. keep the turn beat where it is

you're making ai model videos and they still don't look organic, stop touching the prompt.. it's a finishing problem - i generate everything at 720p, it's cheaper and the upscale comes out way better - 1080p on gpt image or nano banana looks better than 4k to me, 4k adds a weird distortion - upscale once at the end in topaz. starlight precise, about $1.20 per 10 seconds up to 1080p. keeps her face and doesn't invent detail - upscale before you add grain or the colour grade - fix white balance in capcut and pull saturation down. raw generations come out too clean and oversaturated - grain at 2-3% in davinci resolve, plus a pass of jpeg compression artifacts - lay room tone under the voice track - send the finished file through telegram and download it again. one pass of real compression - ig flags pages for ai content, so the final export comes from the instagram edits app, never capcut or davinci. the metadata reads as phone