User
Write something
Getting lip movement right in an H3 character swap
Hi everyone, For my Avengers: Doomsday recast (me as Thor, Hasan as Steve Rogers), the biggest early mistake was speech. Thor stayed silent in a shot where he talks, and in another shot a silent character mouthed words. What fixed it: for shots where someone speaks on screen, I passed the original audio in and wrote the actual spoken words from a transcript into the prompt. For silent shots I said plainly that the character does not speak and keeps the mouth closed. Never write "speaks softly" on a silent shot, the model takes it literally. Built on MDMZ's MiniMax H3 Character Swap workflow in ComfyUI. https://www.youtube.com/watch?v=U85xhNotCe8 Enjoy watching!
Prompts, references and playback: lessons from my local video tests
Hey everyone, a few practical lessons from my Blender-to-ComfyUI tests, including the cafe scene I shared. I hope these save someone some trial and error. 1. Separate the job of each input. Depth helped layout and movement; a real location photo helped preserve the Poznan town hall. Asking for preservation in the prompt alone wasn't enough. 2. Be specific about the cast. Naming clothes and screen positions helped a seven-person corridor shot, though identities and colours still needed checking. 3. Compare models on your actual shot. Wan VACE often held geometry better in my tests; H3 was useful for quick character experiments; LTX sometimes changed people or lighting while improving materials. 4. Keep the source and watch the result. A sharp upscale or good first frame can hide invented details and flicker. Check segment joins in normal-speed playback too. I save the workflow, seed, settings and inputs so the next test has a useful starting point. I'm still learning, and sharing failures feels worthwhile if it helps someone here. Which usually helps you most: a better reference, a clearer prompt, or a different control input?
One generated portrait into a talking cartoon host: what worked and what didn't
Hi everyone, For my latest video I wanted a presenter without filming myself, so I turned a single generated cartoon portrait of me into a talking host. What worked: → Mouth shapes and a blink as edits of the same image, aligned to the original and switched frame by frame from the voice's character timings (cloned voice, ElevenLabs). → A light Blender cutout rig on top: head tilts and nods that follow the voice energy, eyebrow raises at sentence starts, breathing. → Green screen key, then composited into the explainer. What didn't: mouth shapes generated as separate images don't always share the same beard and jaw lines, so wide-open sounds can look off in close-ups. Small, consistent shapes hold up much better than big dramatic ones. The video itself is about running a ~125B Qwen model locally on 16 GB VRAM + 64 GB RAM. https://www.youtube.com/watch?v=pe59kEBAxKk How do you keep mouth shapes consistent when you generate them? Enjoy watching!
Matching real camera views to an AI-assisted Blender location
I've been testing GPT 6.1 Terra inside Blender by rebuilding Poznan's Old Market Square. The most useful shot check came after the build: putting a Blender camera where the reference photographer stood, with the same lens, then wiping between the photo and render. The comparison starts at 1:01 in this 2:47 video. Before that, the screen recordings show the scene growing from OpenStreetMap footprints through roofs and the town hall. I also kept the actual prompts on screen and a section showing geometry mistakes before and after repair. It's an approximate model, with estimated heights and a stand-in fountain, so the matched views help expose what still needs work. https://www.youtube.com/watch?v=_HWXE3wchY4 For those planning shots in 3D, do you match a real reference camera early in the blockout, or wait until the scene's main shapes are in place?
CINEMATIC TWO-CHARACTER LIVE CONCERT VIDEO PROMPT
CINEMATIC TWO-CHARACTER LIVE CONCERT VIDEO PROMPT Create an ultra-realistic, cinematic live concert performance featuring TWO performers on a large professional stage. Use the supplied character reference image(s) as the LOCKED visual reference for both performers. The video should feel like footage captured during a real live concert rather than a music video filmed on an empty set. CHARACTER REFERENCES PERFORMER 1 — [NAME / PERSONALITY] Appearance: [Describe age, hair, eyes, facial characteristics, build, makeup, accessories, etc.] Performance personality: [Reserved / intense / playful / energetic / aggressive / emotional / mysterious / etc.] PERFORMER 2 — [NAME / PERSONALITY] Appearance: [Describe age, hair, eyes, facial characteristics, build, makeup, accessories, etc.] Performance personality: [Describe how this performer behaves differently from Performer 1.] CRITICAL: Preserve both performers' facial identities, hairstyles, body proportions, wardrobe, accessories, and distinguishing features throughout the entire video. Do not merge their faces or identities. If the characters are twins or closely related, maintain their physical resemblance while preserving their individual distinguishing characteristics. WARDROBE Performer 1: [Describe outfit.] Performer 2: [Describe outfit.] Preserve realistic fabric textures, stitching, jewelry, footwear, accessories, and natural clothing movement. No wardrobe changes during the sequence. CONCERT ENVIRONMENT Create a large-scale live concert environment inspired by: [GENRE / VISUAL THEME] Examples: Gothic rock Alternative rock Pop Metal Country Electronic Darkwave Punk Indie Retro Industrial Dream-pop Stage design: [Describe the stage.] Possible elements include: [LED walls / practical lights / theatrical scenery / smoke / lasers / projection screens / pyrotechnics / environmental props / etc.] Include a believable live audience surrounding or facing the stage.
CINEMATIC TWO-CHARACTER LIVE CONCERT VIDEO PROMPT
1-11 of 11
powered by
AIShotStudio
skool.com/aishotstudio-8297
Learn cinematic AI filmmaking: camera moves, looks, prompts, music, voice, and sound. Every prompt and technique here is tested before it is shared.
Build your own community
Bring people together around your passion and get paid.
Powered by