Make a Realistic AI Avatar

A starting recipe for an AI talking-head avatar on Wireflow using a GPT Image 2.5 Sunburst still animated with Seedance 2.5 reference media. Includes observations from individual tests and checks for your own results.

Last updated View as MarkdownAgent setup

This recipe describes one tested talking-head setup, not a guarantee of photorealism or identity preservation. Its observations apply to the reference set and runs described here; other subjects, prompts and model versions can behave differently. Start with a still you approve, then test a short motion take before making a longer cut.

Upload your references first

Upload your reference photos and talking clip to Wireflow and wait for each upload to finish. Add the uploaded media to the canvas, then wire its image or video output to the ports described below. For a mask, upload that image too and wire it to the edit node's mask_url port. A local file path in a prompt does not upload a file or connect a reference.

The still: one GPT Image 2.5 Sunburst pass

Use GPT Image 2.5 Sunburst (generate:openai_gpt_image_2_5_sunburst_text_to_image). Wiring any image into it routes the run to the edit variant (edit:openai_gpt_image_2_5_sunburst_edit) automatically, with the first wired photo landing on image1. That single fact decides how you wire identity refs.

  1. Never let a face photo land on image1. Whatever is wired first becomes the edit canvas, and the model can inherit that photo's pose, light and framing, limiting variation. Two ways to avoid it:
    • No image wired at all: describe the whole shot (subject, pose, outfit, background, lighting) in the prompt and generate from text alone.
    • Identity refs needed: wire a face-free scene or style plate to image1 first, then wire the identity photos into image2, image3 and so on (each additional image you connect mints the next numbered slot). Label each identity ref in the prompt as "identity only, ignore pose, light, expression and clothing," and describe the actual shot in the prompt, not in any of the images.
  2. One face per pass. If two people are in the shot, run two separate passes, each with only that person's refs wired, then composite. Loading both people's refs into one pass can blend their features.
  3. Try Quality max, image_size: custom at 1152x2048 for the first pass. These settings worked for the tested portrait; they do not guarantee a clean result. Inspect the skin, hair and likeness before continuing. Over MCP, update_workflow_config needs (workflows:write OR workflows:create (eligible own draft while mcp_own_draft_edit is on)) and takes image_size as {"width":1152,"height":2048} or "1152x2048"; a bare "custom" is refused. Other sizes and quality settings need their own comparison. Watch for over-sharpened skin, an etched beard or a haloed edge around the hair.
  4. Do not upscale a still that feeds Seedance. An upscale pass can add grain: a Crystal x1.5 pass measured 1.12 to 2.15 on a high-pass grain metric in testing, against 0.37 for a real iPhone photo. Seedance renders at 720p anyway, so a bigger still buys nothing downstream on that path; the pixels get thrown away. A still headed somewhere else, posted as an image in its own right, can still be upscaled the normal way (Crystal Upscaler, creativity 0).
  5. Keep "film grain," "sensor noise" and "pores" out of the prompt. Every old plate prompt asked for grain in words, and paired with an upscale pass that itself added grain, the two together are what made past plates look synthetic. Dropping the wording is one half of the fix; skipping the upscale (rule 4) is the other.
  6. Keep the face large: waist-up or closer, never full body. One test found a face about 40px tall in a 1024x1536 sheet lost identity; published benchmarks separately report similarity dropping from about 0.71 on a large face to about 0.56 on a small one. If just a couple of panels on a sheet come out wrong, aim a masked edit at them (edit:openai_gpt_image_2_5_sunburst_edit, mask_url port). The mask steers where the change lands but does not protect the rest: the model re-renders the whole frame, so grade every panel against the approved sheet after the edit, including panels outside the mask.

The video: Seedance 2.5, reference mode, raw output

Once the still holds up, animate it with Seedance 2.5 (video:seedance_2_5). "Reference mode" is not a setting on the node, it is a wiring pattern:

  • Leave Start Frame (first_frame_url) empty. (image_url is a different port: a plain reference image, not a start frame.) Start Frame cannot be combined with reference media. With Reference Images or Reference Audio the run is refused before you are charged. With a Reference Video wired, the Start Frame is silently turned into one more reference image instead.
  • Wire the approved still, a front crop, a three-quarter crop, and a real mouth close-up into Reference Images (reference_image_urls). Reference images and the Reference Video share one budget of registered materials, and some run paths cap it at 5. Four images plus the talking clip is exactly 5, which is why this version uses one mouth close-up. The set this recipe was tested with used two (6 materials), which those paths refuse before you are charged.
  • For a talking head, also wire a short real clip of the person talking into Reference Video (video_url) as @Video1. The clip guides the generation; its appearance can still influence the output, so check for unwanted carryover. With a Reference Video wired, Wireflow does not send Aspect Ratio, and at the default Duration the take follows the clip's length and orientation. Record the talking clip at the length and orientation you want the take to be. In the tested setup, omitting the real mouth close-up and talking clip led to exaggerated upper teeth and gums. Those references may help, but do not guarantee accurate mouth movement.
  • Wire the still first, and check the @ImageN badge on each port before writing the prompt. The numbering follows wire order, not which slot you meant it for, so wiring out of sequence silently shifts every @Image reference after it.
  • Bind roles in the prompt with @Image and @Video references: ask each to guide a specific aspect of the shot. These prompt instructions are guidance, not enforced boundaries; inspect the whole output for unwanted changes. @Image1 is the composition, wardrobe and background reference. @Image2 / @Image3 are the identity crops, told to "control only the face, do not copy hair or clothing." @Image4 is the mouth close-up, told to control "mouth appearance only." @Video1 is the talking reference, told to control "timing and mouth movement only." Keep mouth wording on the mouth references only. A mouth instruction applied to every reference produced a fixed grimace, teeth bared even while silent.
  • Keep "add," "remove," "replace," "extend" and "continue" out of the prompt. Seedance classifies its task from the prompt text, and any of those words, even inside a "do not copy hair" style instruction, can tip the run into an editing or extension task, which treats @Video1 as the clip to change instead of a reference. Aspect Ratio is not the risk: Wireflow never sends it with a Reference Video.

In the tested takes, this reference setup preserved likeness through a head turn better than first-frame mode (a still wired to Start Frame alone, no reference images). Check your own frontal, profile and turning frames for identity drift.

Ship the raw 720p output. No video upscale by default: of the upscalers tested, some made it visibly worse (a waxy softened look, or an over-sharpened, processed look), and the one that came out closest to native did not do enough to justify the extra credits. The model that scored best on a sharpness or shimmer metric was picked worst in a human side-by-side in one such test. Sharper is not more real. Judge a finish by eye, never by a detail or shimmer score.

Grade and grain, if you want either, go on in the video editor afterward, applied to the whole cut, not baked into the plate or the raw clip.

How to judge the result

Metrics screen out breakage, things like morphing teeth or drifting identity, plus visible shimmer. They do not tell you what looks real, and are not a way to pick a "winner."

  • Always compare side by side, the candidate next to a real photo or a real clip of the same person, not the candidate alone.
  • Re-encode to phone quality before judging. Instagram and TikTok re-encode everything; a clean master can hide grain or artifacts that only show up after a real upload. A candidate carrying 1 percent grain still shows about half of it after an IG re-encode, so judge the compressed version, not the raw export.
  • Trust your eye over the metric. A sharper, more detailed frame reads as processed to a human, which is the opposite of the goal. If the metric's pick and your eye's pick disagree, go with your eye.
  • Check the face at crop zoom, not just at the full frame. Small-frame likeness and crop-zoom likeness are different questions, and only the second one matters once the shot gets cut in tight.

Common mistakes, with the numbers that proved them

Mistake What happened Measured
Upscaling a still that feeds Seedance Grain roughly doubled 1.12 to 2.15 on a high-pass grain metric; a real photo measures 0.37
"Film grain" / "sensor noise" / "pores" in the prompt, plus an upscale pass Old plates written and processed this way looked synthetic Dropping the wording and skipping the upscale, together, fixed it
Picking a finish by a sharpness or shimmer metric In one test, the metric's "winner" was the worst pick by eye A human side by side overruled that metric's pick
Face photo landing on image1 The result looked like a derivative of that one photo's pose and light Fixed by keeping identity refs off image1: a face-free plate there, refs in image2+
Full-body or wide framing Identity fell apart once the face got small About 40px tall in a 1024x1536 sheet lost identity in testing
Two people's identity refs in one pass Features blended between the two subjects Fixed by one face, one pass, composite afterward

When you need a Passport, and when real photos are enough

A Passport is a saved identity reference board: a small set of approved photos (and, for a video model, front plus three-quarter crops) kept with the workflow so every new shot draws from the same refs instead of a fresh photoshoot each time.

Real people with real photos do not need one. If you already have several clear, varied shots of the subject, start with those and check the generated likeness.

Invented characters need a Passport, because there is nothing else to hold identity across shots: generate a reference sheet once, lock it, and reuse it for every future still and clip. A Passport built by GPT Image 2.5 is fine to reuse this way, since grain did not climb pass over pass in testing. Identity does drift a little each time you generate a new sheet from an old one though, so keep drawing from the most recently approved sheet rather than a chain of regenerated ones.

A Passport identifies a real person by face and voice. Keep it private: fork it as a workflow rather than publishing it as a blueprint, and never publish a real person's face or voice as a public blueprint.

Related

  • Clone yourself

    What to record and photograph so Wireflow can make an AI twin of you that sounds like you, looks like you and moves its hands like you. Voice, face, mouth, a talking clip and a gesture clip, in that order, and how each one feeds the tested avatar recipe.

For AI agents: the full documentation index is at /llms.txt, and most docs pages are available as Markdown by adding .md to the URL.

© 2026 Wireflow. All rights reserved.