How to Make a Fashion Lookbook with AI (Full Workflow)

You can produce a full AI fashion lookbook, eight looks, editorial stills plus motion sequences, in about 90 minutes of compute time. The five-step workflow is: garment reference capture, model identity lock with Higgsfield Soul 2.0, scene generation with Seedream 5.0, multi-shot consistency using locked style references, and final cut assembly. Each look costs roughly $25 to $50 in model credits. A comparable traditional editorial shoot runs $5,000 to $25,000 before retouching. The workflow is genuinely production-ready for most fashion categories right now, with a short list of things that still break.

TL;DR

• Best model for editorial stills: Seedream 5.0, best color mood and fabric drape rendering
• Best for model consistency across shots: Higgsfield Soul 2.0 with character reference locking
• Best for motion sequences: Veo 3.1 for cinematic, Reve for stylized/artistic
• Cost per look: $25 to $50 in compute vs $5K to $25K traditional editorial
• What still fails: fine lace and silk textures drift, repeated patterns shift across cuts, logo embroidery isn't accurate

Why fashion is a sweet spot for AI right now

Three things converged in early 2026 that make fashion the clearest production use case for AI generation.

First, consistency across cuts is solved for human subjects. Higgsfield Soul 2.0's character reference conditioning holds a model's face, skin tone, and general proportions across 10 to 15 shots before drift becomes noticeable. That was the blocking problem for fashion brands through most of 2024 and 2025. A lookbook where the model looks like three different people is unusable.

Second, garment rendering improved substantially in Q1 2026. Seedream 5.0 added layered cloth simulation to its conditioning stack. You'll see the difference most on structured pieces like blazers and outerwear: the garment sits on the body instead of floating near it. Soft knits and tailored trousers now read as fabric. The gap from earlier models is large enough that workflows built before Q1 2026 need revisiting.

Third, scenes are reusable across drops. Once you've generated a clean urban rooftop scene or a studio-white backdrop in a consistent lighting setup, you can reuse that scene file for every new season. The scene becomes a brand asset. Traditional photography sets don't work that way.

Fashion editorial was always about controlled surfaces: controlled lighting, controlled sets, controlled looks. AI generation has the same constraint profile. That's why it works better here than in, say, documentary or live event coverage where unpredictability is the point.

The 5-step AI fashion lookbook workflow

Step 1: Garment reference

This step determines everything downstream. The model needs to know what the garment looks like before it can generate it.

You have two options. The first is real photos: flat lays, product shots on a white background, or even phone photos of the garment on a hanger. Seedream 5.0 and Higgsfield handle reference images well as long as the photo is sharp and the garment is the dominant subject. Take 2 to 3 references per piece: front, back, and a detail shot for anything with surface texture.

The second option is fully generated garments using Nano Banana Pro for product detail work. If you're working on a not-yet-produced collection or want to concept new colorways, you can describe and generate the garment first, then use that generated image as your reference input downstream. Nano Banana Pro's render quality on fabric structure is the best of the image models we've tested for this step.

Keep one rule: each garment gets its own reference file. Don't combine multiple pieces in a single reference image. Models interpret them as one outfit and you lose garment-level control.

Step 2: Model identity lock (Higgsfield Soul 2.0)

Open Higgsfield Soul 2.0 in your 8frame canvas. Upload a character reference: either a real photo of a model you've licensed rights to, or a generated character from Seedream or Nano Banana. This reference becomes the identity anchor for every shot in the campaign.

The key parameter is reference weight. Start at 0.75 to 0.85. Lower weights give the model more compositional freedom but drift identity faster. Higher weights preserve the face closely but sometimes constrain pose and expression range. For a 10-shot lookbook, 0.80 is the working default.

Run a test batch of 3 to 4 shots before committing the full campaign. Look specifically at the jawline, eye distance, and skin tone across shots. If those three hold, the identity is locked well enough for editorial use. If they're drifting, raise the reference weight by 0.05 and retest.

Don't try to get the clothing right at this step. You're locking the person, not the outfit. The garment comes in at step 3.

Step 3: Scene generation (Seedream 5.0 for stills, Veo 3.1 for motion)

With the identity locked, you now generate the look. Add the garment reference from step 1 alongside the character reference in your 8frame canvas. Seedream 5.0 handles the compositing.

Prompt structure for editorial stills: setting first, lighting second, pose third. An example from an indie streetwear campaign:

Rooftop in lower Manhattan, late afternoon light, slight overcast. Woman in oversized olive utility jacket, dark trousers. Standing, looking off-frame left. Editorial fashion photography. Canon R5 aesthetic. No text.

Generate 6 to 8 variants per look at this stage. You'll keep 2 to 3. Seedream's color moods are strong: golden hour looks like golden hour, studio strobe looks like studio strobe. That mood consistency across variants is why it's the pick for editorial stills over other image models.

For motion sequences, switch to Veo 3.1. Take your best Seedream still, use it as the first-frame reference, and prompt a short motion: a slow walking shot, a jacket brushing in wind, a coat turn. Veo handles the cloth physics on motion better than any other video model in this workflow. Budget roughly $0.85 to $1.20 per 5-second motion clip.

If the brief calls for a more artistic, non-photoreal treatment, Reve is the swap for Veo. Reve's stylized rendering is good on fashion: it abstracts toward illustration or high-contrast editorial photography depending on the prompt. Use it when the lookbook has an intentional non-documentary aesthetic.

For a deeper look at cinematic transitions between motion clips, the dissolve-appear workflow on 8frame is directly applicable here.

Step 4: Multi-shot consistency

A lookbook is not a single image. It's a sequence, and every shot needs to read as the same campaign.

Consistency across shots comes from three locked elements: the character reference (same Higgsfield identity file throughout), the style reference (a mood board of 2 to 3 images that establish the campaign's visual language, loaded into Seedream as a style reference), and the scene reference (if using a specific setting across multiple looks, generate it once and use that generation as the background reference for subsequent shots).

The workflow in 8frame is: one canvas, all 8 looks, each look as a separate node chain. The character and style references wire in at the top and feed every chain. This makes it easy to update the reference globally if a shot drifts and you need to rerun the set.

One honest note: Higgsfield's identity lock holds well for 10 to 12 shots. Beyond that, in our experience, faces start to drift subtly. For a 20+ look campaign, plan to run two generation batches with a consistency check between them, not one continuous run.

For motion sequences specifically, using the first frame of each Seedream still as the Veo input keeps the visual continuity tight. The Christmas editorial guide walks through a high-fashion scene setup that demonstrates this first-frame-to-motion technique in a different context.

Step 5: Final cut

Assemble the stills and motion clips in your editing timeline. The 8frame canvas exports each output as a separate file, so you're working with clean individual assets.

Three things that matter at this stage:

Color grade should be consistent across all clips. Even though Seedream is good on mood, individual clips will have minor exposure differences. A single LUT applied across the set fixes this in 20 minutes. Don't skip it.

Transitions between motion clips can make or break the editorial feel. The dissolve-appear transition workflow on 8frame handles this natively if you want a specific kind of cut.

Music affects how loose the editing can be. A slower, textural track covers cuts that would read as choppy at a faster tempo. Match the music to the pace of the motion sequences you generated, not the other way around.

For an 8-look campaign with one motion sequence per look plus a cut-down 60-second edit, final assembly typically runs 3 to 4 hours for someone who knows the software. The generation itself is the 90-minute part.

Sample walkthrough: indie streetwear drop, 8 looks

Here's the actual process for a recent streetwear campaign that went from brief to final assets in one afternoon.

The brief: eight looks from a new collection, urban environment, late afternoon light, one recurring model, demographic 18 to 28. Budget: $400 for the whole campaign.

Garment reference: Product shots on white background for all 8 looks, taken on an iPhone with a light tent. 16 total reference images (front and back for each piece).

Identity lock: Generated a base character in Nano Banana Pro using a written description matching the brand's aesthetic. Uploaded that to Higgsfield Soul 2.0 as the character reference at weight 0.82.

Scene generation: Lower Manhattan rooftop for looks 1 to 4, warehouse alley for looks 5 to 8. Generated one clean scene base per location, then used those as background references. Each look went through Seedream 5.0 at 6 to 8 variants; we kept 2 selects per look.

Motion sequences: Veo 3.1 on 3 of the 8 looks for a short motion clip each (5 seconds). Total spend on Veo: $12.40.

Deliverables: 16 editorial stills (2 per look), 3 motion sequences, one 60-second cut-down, one 15-second Reels cut.

Total compute cost: $183. Generation time: 94 minutes wall-clock. Final assembly and grading: 3 hours.

Comparable traditional shoot in New York: $8,000 to $14,000 for location, photographer, model, stylist, and retouching. Two-week turnaround minimum.

Where it falls apart

This workflow has real limits. Don't oversell it to a client and find out mid-project.

Fine fabric textures. Lace, silk charmeuse, and sheer organza are still uncanny. The models don't accurately render the way light passes through translucent fabrics. You'll get something that looks like fabric in a screenshot and looks wrong in playback. For collections built around these materials, AI generation is still a concept tool, not a campaign tool.

Repeat patterns. A specific plaid, a stripe repeat, a branded jacquard: these drift across shots. If look 1 has 3cm plaid and look 4 has 4cm plaid, that's going to read wrong to anyone who knows the collection. The garment silhouette holds, the pattern doesn't always. Test your specific fabrics before committing to a full run.

Brand logo accuracy. If the garment has a printed logo, embroidered text, or any typographic element, assume it will not be accurate. AI models generate plausible-looking logos, not correct ones. Plan to either remove logos from the reference images and add them in post, or work with looks that don't feature prominent branding elements.

Model likeness. If you want to generate a specific real person, you need their explicit consent and a licensing agreement. This applies to celebrities, brand ambassadors, and anyone recognizable. Generating a fictional character from scratch has different considerations, but they're not zero.

Pricing math vs traditional editorial

Specific numbers from our campaign experience in Q2 2026:

Item	AI lookbook workflow	Traditional editorial shoot
8-look still campaign	$150 to $300 compute	$8,000 to $25,000
Per-look cost	$19 to $37	$1,000 to $3,000
Motion clips (3 x 5s)	$10 to $15	$2,000 to $8,000
Turnaround	Same day	1 to 3 weeks
Revisions	Rerun for $5 to $20	Reshoot negotiation

The cost difference is real. The quality difference is also real on the failure cases above. The honest framing for clients: AI fashion editorial is production-ready for collections without fine sheer fabrics, without critical repeat patterns, and without strict logo requirements. That covers a significant portion of streetwear, outerwear, knitwear, denim, and suiting.

Models compared for fashion work

Every model in this workflow has a specific role. Here's the breakdown for fashion specifically.

Seedream 5.0 is the editorial still workhorse. Color science is accurate, mood is consistent, fabric drape reads well on structured garments. Use it for 80% of still generation in a lookbook workflow.

Nano Banana Pro is for product detail and garment generation. If you need a reference image that shows a specific surface texture, button configuration, or seam detail, this is the model. Sharper on fabric specifics than Seedream at small detail scale.

Higgsfield Soul 2.0 handles the model identity lock. This is its singular job in the workflow. Don't use it for scenes or garment rendering. The character reference system is the best available for multi-shot consistency in fashion.

Veo 3.1 is for motion sequences when cinematic quality is the brief. Cloth physics on motion, lighting coherence, depth of field in motion: Veo outperforms everything else. The cost is real ($0.85 to $1.20 per 5s clip) but a fashion brand can absorb that on hero motion assets.

Reve is for stylized or artistic campaigns where photorealism isn't the goal. High-contrast editorial, illustrative lookbooks, or campaigns intentionally leaning away from a documentary aesthetic. Reve handles those better than pushing Seedream or Veo into non-photoreal territory.

More on how these models compare across other use cases: best AI video generator 2026 and all 8frame workflows.

A note on legal considerations

Commercial use of AI-generated fashion content involves three areas to get right. Model likeness: if you're generating a realistic human character, that character needs to be clearly fictional or you need documented consent from any real person whose likeness was used as input. Garment IP: generating an existing designer's piece doesn't make the output licensable. If the reference image is a Loro Piana coat, the output is still derivative. Use AI generation for original designs or clearly modified interpretations. Brand consistency: if a generated output is going to represent a licensed brand, the brand owner will want approval rights over the output. Build that review step into your workflow before final delivery.

FAQ

Can an AI model look like a real person?

It can, but using a real person's likeness as input without explicit consent and a licensing agreement is a legal exposure. Generate fictional characters, or if working with a real brand ambassador, get the consent documentation in place first. Most clients will ask about this.

How do I keep the garment looking the same across shots?

Two things. First, use a clean, dedicated reference image for each garment (not combined with other pieces). Second, in Seedream 5.0 and Higgsfield, keep your prompt consistent: same garment description language, same lighting descriptors. The model doesn't read your previous outputs; it reads the reference and the prompt. Consistency in inputs produces consistency in outputs.

What's the best aspect ratio for fashion editorial AI content?

3:4 or 4:5 for editorial stills (portrait orientation, closer to print magazine proportions). 9:16 for Reels and TikTok motion content. 16:9 if you're cutting a campaign video for YouTube or a brand site header. Generate at the delivery ratio; cropping after the fact loses quality and sometimes loses the composition.

Can I use AI fashion content commercially?

Yes, on 8frame's paid tiers. Every model accessed through 8frame's paid plan permits commercial use of the output. Check the model card for any model-specific restrictions. Free tier outputs vary by model: some permit commercial use, some don't. Don't ship client work on the free tier without checking.

How do I handle fabric textures like silk?

Plan around them rather than through them. For silk, satin, and other specular fabrics: generate the look, then evaluate the fabric rendering specifically before committing to a full run. If it's reading wrong (flat, plastic-looking, or with incorrect specularity), either rewrite the prompt to describe the material more specifically ("heavy charmeuse, catching direct window light, slight sheen") or use the AI output as a composition reference and re-shoot or retouch the fabric areas. Lace remains the hardest category. Current models don't accurately render lace transparency. Treat it as a manual compositing job rather than a generation job.

Run the fashion lookbook workflow

All five steps in this article run on 8frame's multi-model canvas. You can route garment reference through Nano Banana, lock identity in Higgsfield, generate stills in Seedream, and cut motion in Veo inside a single session. Browse the full 8frame workflow library to see starting templates, or start from scratch with your first garment reference.