A brand consistency AI workflow solves one problem: when you chain several models across a campaign, each one reinterprets the brief, and the set looks like it came from four studios. The fix is to lock a visual reference before any model sees a prompt, feed that reference into every node, and make every video from an approved on-brand still. On 8frame, a 22-asset, 4-format campaign built this way costs about 1,700 credits.
Correction, 2026-09-30: the first version of this page described a multi-format export node and a colour-check node that measures drift against a tolerance (neither exists on 8frame), reported a campaign run with timings and a measured colour spread we can't reproduce, generated 4:5 stills on Seedream (Seedream 5 has no 4:5; Nano Banana does), listed 4K stills from Seedream (it tops out at 3K), and described fixed colour tendencies per model that we can't source. This version keeps the method and uses the canvas as it is.
TL;DR
- Drift is a chaining problem. Fix it with one locked reference image in every node, not longer prompts.
- Five steps: reference lock, palette fragment, per-format composition fragments, a still per shape with video animated from it, and a side-by-side review.
- 22 assets in 4 formats: about 1,700 credits, including a few re-runs.
- Keep text and logos out of generations; add them in post.
Why models drift when you chain them
Each model renders colour, contrast and light its own way. None of that is wrong alone, but run a campaign through three or four models without a shared anchor and the outputs land in different places: the hero still looks like autumn, the social cut like a different brand. Writing "warm neutral tones, 3200K" in every prompt helps a little; each model still reads it differently. The reliable fix is a locked visual reference fed into every model, before the text.
The 5-step workflow
Step 1: Build the brand-locked reference image
One clean still that shows the exact palette, lighting angle, background treatment and product placement you want. Not a mood board: one image. Build it on Nano Banana Pro (21 credits, 42 at 4K), with your values in the prompt:
Background warm grey #E8E4DC, product lit from upper-left 45 degrees, no fill light on right side, gold accent light on the product edge, product centered, no text
Or start from real brand photography: a real product photo gives the most accurate colour and surface for the product itself. Either way, connect this image as a reference to every generation node in the campaign.
Step 2: A palette fragment in every prompt
The reference constrains the model visually; the palette fragment constrains it in words. Both together are stronger than either alone.
Brand palette: primary #1A1A2E (deep navy), accent #C9A84C (warm gold), neutral background #F5F1EB (warm off-white). No cool greys, no blue-tinted shadows. All highlights warm.
Keep it in your brief and paste it at the start of every prompt.
Step 3: A composition fragment per format
Each format has its own framing logic. Stack one of these after the palette fragment:
- Still, 4:5 editorial: "Tight composition, subject fills 70% of frame, shallow depth of field, editorial grade"
- Still, 16:9 web: "Wide composition, environmental context visible, medium depth of field, clean background"
- Video, 9:16 social: "Vertical framing, subject enters from bottom third, energetic but smooth motion"
- Video, 16:9 brand: "Cinematic framing, slow camera movement, no rapid cuts"
Step 4: A still per shape, video from the still
Generate each shape natively rather than cropping, and pick models that offer it:
- 4:5 stills: Nano Banana (Seedream 5 has no 4:5).
- 16:9 stills and 9:16 start frames: Seedream 5 Lite, 6 credits each, up to 14 references.
- 9:16 social clips: Kling v3 from the approved 9:16 still (it follows the start image's shape), 57 credits for 5 silent seconds.
- 16:9 brand clips: Veo 3.1 from the approved 16:9 still, 8 seconds, 224 credits silent.
Animating approved stills, instead of prompting video from text, is what carries the palette into motion. Name each file by campaign, format and number as you download (cosm-launch-social-03), so the teams downstream don't have to open files to find the right one.
Step 5: Review side by side
Put every output on the board next to the reference and look at them together: background tone, shadow colour, highlight warmth, product colour. When one drifts (a cool cast in the shadows is common), re-run it with the fix in its format fragment ("warm shadows, no cool blue in background shadow areas") rather than correcting it in post.
Walkthrough: 22 assets, 4 formats, costed
A cosmetics launch with one hero product: 6 editorial stills (4:5), 4 web stills (16:9), 8 social clips (9:16) and 4 brand clips (16:9).
| Item | Model | Count | Credits |
|---|---|---|---|
| Reference still, 4K | Nano Banana Pro | 1 | 42 |
| Editorial stills, 4:5 | Nano Banana Pro | 6 | 126 |
| Web stills, 16:9 (also brand clip start frames) | Seedream 5 Lite | 4 | 24 |
| Social start frames, 9:16 | Seedream 5 Lite | 8 | 48 |
| Social clips, 5 s silent | Kling v3 Standard | 8 | 456 |
| Brand clips, 8 s silent | Veo 3.1 | 4 | 896 |
| Re-runs for drift | Kling v3 Standard | 2 | 114 |
| Total | 1,706, about $17 |
With Veo 3.1 Fast for the brand clips (112 each silent) the total is 1,258. For stills in more depth, see Seedream 5 prompts for brand visuals.
Pitfalls
Model colour habits. Models differ in how they render shadows, contrast and highlights. You'll see it when the set is side by side; compensate in that format's fragment.
Fonts. Models don't have your typeface. Text in a generation comes out as a lookalike that anyone who knows the brand will spot. Keep type out and add it in post.
Logos. Same problem, higher stakes: models treat a logo as a texture and reshape it. Composite the real logo after review.
FAQ
Can I use existing brand photography as the reference?
Yes. A real product photo gives accurate colour and surface from a real capture; a generated reference built to spec gives more control over lighting and background. For brands with strong photography, starting from a real photo usually gets the product colour closest.
How many formats can one workflow handle?
Four to six before review gets slower than it's worth. Beyond that, split into two boards that share the same reference image.
What if one format keeps drifting?
That model has a habit your fragment isn't overriding. Adjust the fragment for that format rather than tightening the global palette instruction, and check the start frame: a clip can't be warmer than the still it starts from.
Sources
- 8frame Nano Banana, Seedream, Kling and Veo prices, shapes and reference limits: the 8frame tool registry and aspect-ratio configuration, 2026-09-30
For agency setups, see AI for marketing agencies; starting points are in the workflow library. The 8frame canvas is free and unlimited; generation is paid from $19/month.