Seedream 4.5 Complete Guide: ByteDance's Best AI Image Model

Seedream 4.5 is useful when an image needs more than a convincing subject. A poster also needs readable copy; a product campaign needs the same bottle to look like the same bottle after a background change. ByteDance highlights typography, dense text, and reference-image consistency as improvements in 4.5. This guide covers those strengths, their limits, and the generation and editing paths available through WaveSpeedAI.

“Best” depends on the job. The original draft described Seedream 4.5 as a current leaderboard leader, but its rank and score had no dated source. The practical question is whether 4.5 produces an image you can use with less correction than the alternatives you are considering.

Introduction to Seedream 4.5

Seedream 4.5 is a ByteDance image model for text-to-image generation and reference-guided editing. ByteDance’s Seedream 4.5 overview emphasizes improved typography, dense text, and preservation of subjects and details during multi-image editing.

It is available through more than one provider. This guide uses WaveSpeedAI for its API examples; fal’s text-to-image endpoint and editing endpoint also list version 4.5. Provider endpoints and parameter names are different, so keep the examples below with the provider they were written for.

Key Statistics

• Version covered: Seedream 4.5, rather than the newer Seedream 5 family
• Highlighted capabilities: Typography, poster composition, and reference-guided editing
• WaveSpeedAI text-to-image default size: 2048*2048
• WaveSpeedAI editing inputs: An images array accepting up to 10 reference images in the current schema
• WaveSpeedAI listed text-to-image price, checked October 30, 2026: $0.04 per generated image; confirm the live price before budgeting

A 2048×2048 image has about 4.2 million pixels. It should not be described as “4K” solely because of that pixel count. Available dimensions also depend on the endpoint’s current schema.

What’s New in Version 4.5

ByteDance describes 4.5 as an improvement over 4.0 in prompt adherence, visual quality, typography, and editing consistency. Its published comparisons are ByteDance’s own evaluation, so they are a useful description of what the release targets, rather than an independent guarantee for every prompt.

Enhanced Text Rendering Engine

The most relevant change for posters and brand visuals is stronger handling of:

• Multi-line layouts
• Titles and smaller supporting copy
• Text placed within a scene
• Dense text in designed compositions

That does not guarantee correct spelling, punctuation, or figures in every output. Proofread any generated copy before publishing, especially prices, ingredients, dates, and legal text.

Improved Image Quality

Version 4.5 aims to preserve subject details while improving composition and image fidelity. That matters most when a visually attractive result still has to retain a specific product shape, label, or reference subject. Inspect those details at the final delivery size; a strong-looking thumbnail can hide an altered logo or an incorrect line of text.

Advanced Style Control

Prompts can guide lighting, palette, composition, and artistic treatment. For an existing asset, use the editing path and say both what should change and what must remain. For a new asset with no source image, use text-to-image generation.

Key Features Breakdown

Professional-Grade Output Quality

Seedream 4.5 can be used for:

• Posters and campaign visuals
• Social graphics
• Product imagery
• Presentation and landing-page illustrations
• Editorial-style artwork

“Professional-grade” describes a potential use, not an automatic approval for print or publication. Check dimensions, copy, brand colors, and product details against your production requirements.

Prompt Understanding

Detailed prompts are useful when they assign clear roles to the subject, layout, and text. For example, “white bottle centered, dark green background, title at the top” gives the model a clearer task than “make a premium ad.”

The WaveSpeedAI 4.5 text-to-image schema documents prompt and optional size. It does not list negative_prompt, quality, n, or seed for that endpoint. Describe exclusions in the prompt where useful, and check the specific endpoint schema before adding parameters from another image API.

Versatile Application Support

The same model family serves different tasks, but the endpoint matters:

• Text-to-image: Create an image from a prompt.
• Edit: Supply reference images and describe a change.
• Sequential: Request a related set of generated images.
• Edit sequential: Work on a related edited set.

Choose the task first. Repeated calls to the single-image endpoint do not, by themselves, lock a character or product across a campaign.

Text and Typography Rendering

Typography is one of the reasons to consider Seedream 4.5. ByteDance shows poster layouts and text-heavy visual examples, including product displays. The useful test is whether your exact required copy survives generation.

Text Rendering Capabilities

Short display copy: Titles, signs, and prominent labels are sensible starting points.

Font direction: You can request “bold sans-serif,” “handwritten,” or “serif,” although a prompt is not a reliable way to specify an exact licensed font family.

Text integration: Copy can be requested on signage, packaging, posters, or within a scene. Keep the intended location explicit.

Best Practices for Text Generation

1. Quote the required words. Example: A blue sign reading "OPEN 24 HOURS" in large white letters.
2. Describe placement. Example: Put "The Daily Grind" on the wooden storefront sign, above the door.
3. Separate text from style. Example: Headline: "INNOVATION"; bold black sans-serif; centered near the top; white background.
4. Check the result character by character. A convincing layout does not prove the copy is correct.

Limitations and Workarounds

Long paragraphs, tiny type, mixed scripts, and exact product copy need closer review. Generate several candidates if the layout is flexible. If a price, ingredient list, or legal statement must be exact, add or correct that text in a design editor after generating the visual. For a small correction to an existing image, try the edit endpoint with a narrowly scoped instruction before regenerating the entire design.

Image Quality and Resolution

The original draft called 2048×2048 both the maximum and “4K quality.” Neither is a safe general rule. WaveSpeedAI currently lists 2048*2048 as the text-to-image default and gives larger suggested dimensions, subject to its live input schema. The dimensions that work on one provider should not be copied unchanged into another provider’s request.

Resolution Options

WaveSpeedAI’s 4.5 documentation gives these examples:

• Square: 2048*2048
• Landscape 16:9: 2560*1440
• Landscape 4:3: 2688*2016
• Large square example: 4096*4096

These are examples, not a promise that every custom width and height will be accepted. Check the current schema when choosing a custom size.

Quality Characteristics

Review outputs for:

• Sharpness at the intended display size
• Edges around products, hands, and text
• Smooth gradients and believable lighting
• Correct copy and preserved reference details

For print, calculate the required pixel dimensions from the physical size and print resolution. A square image that looks sharp on a phone may still be too small for a large-format placement.

Optimization Tips

1. Match size to delivery. Start with a documented size and increase it only when the layout needs more pixels.
2. Describe visible qualities. “Soft studio light from the left” gives clearer direction than a stack of “high quality” adjectives.
3. Review the actual file. Zoom in on text and logos before choosing a candidate.
4. Change one important variable at a time. If both layout and copy are wrong, fix the prompt’s hierarchy before trying a larger output.

Multi-Image Consistency

A consistent palette is easier to request than a perfectly unchanged product. Treat those as separate checks.

Consistency Features

For a related series, prompts can repeat:

• Color palette
• Lighting direction
• Camera angle
• Background treatment
• Subject description

When an exact subject must carry across images, provide references through an editing workflow. WaveSpeedAI also lists a separate sequential endpoint for generating related outputs in one request.

Techniques for Multi-Image Projects

1. Write a fixed visual brief. Keep the product name, color, materials, lighting, and framing terms stable.
2. Change only the intended variable. For example, vary the background while keeping the bottle and label fixed.
3. Use the appropriate endpoint. A prompt-only series and a reference-guided edit are different tasks.
4. Save the inputs and outputs. Keep the prompt, reference URLs, endpoint, size, and selected result for each asset.

The original advice to use “similar seed ranges” was removed because the documented WaveSpeedAI 4.5 text-to-image parameters do not include a seed field. A seed, even where supported, would not establish brand consistency by itself.

Use Cases

Multi-image consistency matters for social campaigns, presentation decks, storyboards, brand assets, and product showcases. Decide in advance what “consistent” means for the job: matching colors may be enough for an illustration series, while a product campaign may require the label, cap, proportions, and colorway to remain unchanged.

Product-Consistent Editing: A Practical Check

Suppose you have an approved photo of a skincare bottle and need a second campaign image with a different background. Start with the approved photo as the reference. In the 4.5 edit request, supply its accessible image URL in images and use a focused prompt:

Replace the background with a pale green studio setting. Keep the bottle shape, cap, label wording, logo placement, product color, and camera angle unchanged. Match the original light direction and add a natural shadow.

Before accepting the output, compare it with the source at full size. Check the label letter by letter, the cap and bottle proportions, the edge of the product, and the direction of the shadow. Record each failure separately: “label changed” is more actionable than “doesn’t look right.”

For several campaign frames, decide whether you need one edit per approved composition or a related set through the edit-sequential path. Do not assume that placing several URLs in images means you will receive one independently edited output for each input; inspect the endpoint’s returned outputs and its current schema. This example is a review procedure, not a claim that a particular test run passed.

Style Transformation Capabilities

Seedream 4.5 can generate an image in a requested visual style. With a reference image, the edit path can also change the look of an existing asset. Those operations should not be treated as identical when exact subject preservation matters.

Supported Style Categories

Photographic directions: Portrait, landscape, product, cinematic, or documentary-style imagery.

Artistic directions: Oil painting, watercolor, digital illustration, anime, comic art, or line drawing.

Design directions: Minimalist, vintage, futuristic, art deco, Bauhaus, or isometric layouts.

Specialty directions: Low-poly 3D, pixel art, paper cut-out, stained glass, or neon lighting.

These are prompt directions, not a fixed menu of guaranteed styles.

Style Prompting Techniques

Direct style description:

A mountain landscape painted with loose impressionist brushwork.
A robot character in a colorful anime illustration.

Combined directions:

A cityscape with cyberpunk lighting and traditional Japanese print textures.
A product photograph with minimalist composition and vintage color grading.

Technical direction:

Studio portrait, soft light, shallow depth of field, 85 mm lens perspective.
Architectural scene at golden hour, long shadows, eye-level view.

Style Consistency Tips

Keep the palette, lighting, and composition terms stable across related prompts. If an existing product or person must remain recognizable, use a reference image and explicitly list the details to preserve. Review the returned images rather than relying on prompt wording alone.

API Usage via WaveSpeedAI

WaveSpeedAI provides a documented REST route for Seedream 4.5. The following examples use its current model identifier and asynchronous prediction flow, as shown in the WaveSpeedAI Seedream 4.5 API documentation.

Getting Started

1. Create a WaveSpeedAI account and obtain an API key.
2. Confirm that your chosen endpoint is text-to-image, edit, or sequential.
3. Start with one documented size and a short prompt.
4. Save the returned prediction ID so you can retrieve the result.

As checked on October 30, 2026, WaveSpeedAI listed $0.04 per generated image for its 4.5 text-to-image endpoint. Check the live listing and billing terms before estimating a large batch.

API Authentication

Send the key in the Authorization header:

Authorization: Bearer YOUR_WAVESPEED_API_KEY
Content-Type: application/json

Keep the key in an environment variable or server-side secret store. Do not put it in browser code or publish it with an example.

Basic API Request Structure

For a text-to-image task, submit the request to the model endpoint:

curl -X POST \
  "https://api.wavespeed.ai/api/v3/bytedance/seedream-v4.5" \
  -H "Authorization: Bearer $WAVESPEED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt":"A minimalist coffee poster with the title \"THE DAILY GRIND\" at the top","size":"2048*2048"}'

The model identifier is part of the URL. It is not a required model field in this request body.

API Parameters

For this WaveSpeedAI text-to-image endpoint:

• Required: prompt
• Optional: size, written as width*height; the documented default is 2048*2048
• Output options: The current schema also lists options such as base64 output and synchronous mode

For ​editing​, use POST /api/v3/bytedance/seedream-v4.5/edit with prompt and an images array. WaveSpeedAI currently documents up to 10 reference images for that endpoint. For a related generated set, use the separate sequential endpoint and its max_images parameter. Do not substitute n, quality, or negative_prompt from the original draft into these requests without checking the live schema.

Response Format

A submission returns a prediction object containing an ID. Query:

GET https://api.wavespeed.ai/api/v3/predictions/{id}/result

When data.status is completed, read data.outputs. An output is commonly a URL under the default output setting. While the task is pending, outputs may be empty. This differs from the immediate data[0].url response shown in the original draft.

Error Handling

• Invalid request: Check the endpoint, JSON field names, size format, and reference-image URLs.
• Authentication failure: Check that the bearer key is present and valid.
• Rate limit or temporary polling failure: Wait and poll the existing prediction ID again.
• Task failure: Inspect data.status and data.error.

Do not automatically resubmit a generation request after an uncertain network failure: the first request may have created a billable task. Check your prediction history or existing task ID first.

Code Examples

The examples below use Python’s requests package. Install it with pip install requests, set WAVESPEED_API_KEY, and keep the helper in the same script as the examples that call it.

Python Example

import os
import time
import requests

API = "https://api.wavespeed.ai/api/v3"
HEADERS = {
    "Authorization": f"Bearer {os.environ['WAVESPEED_API_KEY']}",
    "Content-Type": "application/json",
}

def generate(prompt, size="2048*2048"):
    response = requests.post(
        f"{API}/bytedance/seedream-v4.5",
        headers=HEADERS,
        json={"prompt": prompt, "size": size},
        timeout=60,
    )
    response.raise_for_status()
    task_id = response.json()["data"]["id"]

    while True:
        result = requests.get(
            f"{API}/predictions/{task_id}/result",
            headers=HEADERS,
            timeout=60,
        )
        result.raise_for_status()
        data = result.json()["data"]

        if data["status"] == "completed":
            return data["outputs"]
        if data["status"] in {"failed", "cancelled", "timeout", "deleted"}:
            raise RuntimeError(data.get("error") or data["status"])
        time.sleep(2)

outputs = generate(
    'A studio product photo of a smartwatch, white background, '
    'with "TIME TO INNOVATE" on the display'
)
print(outputs)

The loop polls an existing task. It does not retry the submission. For a production application, add an overall polling deadline and log the task ID before waiting.

Python Example (Alternative)

Using the generate function above:

outputs = generate(
    'A vibrant street-art mural with "CREATE YOUR FUTURE" '
    "in large graffiti lettering on an urban wall"
)
print(outputs)

Use this second prompt to check both the visual style and the exact letters. If the copy is wrong, refine the instruction or correct the final artwork in an editor.

Python Example with Text

outputs = generate(
    "A warm, photorealistic coffee shop interior. "
    'On the chalkboard, write exactly: "DAILY SPECIALS". '
    "Keep the heading large and leave the smaller menu area blank."
)
print(outputs)

Leaving the detailed menu blank makes sense when prices must be exact. Add them in a design tool after approving the image.

Batch Generation Example (Python)

prompts = [
    'A poster titled "MONDAY MOTIVATION" over a sunrise',
    'A clean workspace with "PRODUCTIVITY TIPS" on a laptop screen',
    'A coffee campaign image with "FUEL YOUR DAY" above the cup',
]

for index, prompt in enumerate(prompts, 1):
    outputs = generate(prompt)
    print(f"{index}/{len(prompts)}: {outputs}")

This is three separate text-to-image requests. It does not create a reference-locked series, and each successful generation may be billed separately.

Comparison with Competitors

A fair comparison uses the same deliverable, not a table of unsupported ranks. Give each model the same exact headline, reference asset when editing is involved, aspect ratio, and review criteria. Record the provider and model version as well as the result.

Seedream 4.5 vs DALL-E 3

For a typography-heavy brief, compare the number of corrections needed to make the headline publishable. Also compare whether the image layout leaves room for final copy. The original draft claimed a specific ranking advantage over DALL-E 3 without a dated leaderboard snapshot; that claim should not guide a current decision.

Seedream 4.5 vs Midjourney

If the deliverable is a sequence of brand assets, compare repeatability and the workflow available to your team. If the deliverable is a single expressive image, compare the actual visual choices. Avoid assuming an access method or feature set is unchanged from an older Midjourney version.

Seedream 4.5 vs Stable Diffusion

The practical distinction for many teams is managed API work versus a self-managed model workflow. Compare setup time, control, recurring cost, and the effort required to keep the same output process available to the team. “Stable Diffusion” covers multiple versions and implementations, so name the one you test.

Seedream 4.5 vs Flux

Compare your chosen Flux version with 4.5 on the same prompt and reference material. For product work, score preserved labels, shape, and material; for posters, score text accuracy and composition. A general claim that either family is better at text “in most cases” needs a dated, reproducible test.

Comparison Criteria Summary

Criterion	Use the same input	Check the output
Typography	Exact headline and placement	Every character, line break, and reading order
Product editing	Same approved reference image	Label, logo, shape, colorway, and edges
Series consistency	Same visual brief and number of assets	Recurring subject and palette
Delivery	Same target size and format	Usable dimensions and correction time
Cost	Same number of accepted images	Total spend, including rejected generations

This table replaces the original rank-and-score comparison because the figures were not tied to a date, leaderboard category, or source record.

Best Practices and Prompting Tips

The most useful prompt describes the visible result and identifies the details that cannot change.

Effective Prompt Structure

A workable starting structure is:

[Subject] + [Setting] + [Composition] + [Style] + [Exact text, if any] + [Details to preserve, if editing]

Example:

A speaker on stage in a modern auditorium, viewed from the audience.
A large slide behind the speaker reads exactly "Q4 RESULTS".
Dark blue background, white sans-serif headline, cinematic lighting.

Prompting Best Practices

5. Be specific and descriptive

• Vague: “A nice poster”
• Better: “A minimalist movie poster, dark blue background, large white title at the top”

6. Use style words that affect the image

• Photography: “studio lighting,” “shallow depth of field,” “golden-hour light”
• Art: “watercolor,” “oil painting,” “line drawing”
• Design: “minimalist,” “art deco,” “geometric layout”

7. Specify text clearly

Quote the required words, state where they belong, and review the output manually.

8. Control composition

State the viewpoint, framing, focal subject, and space reserved for copy.

9. Manage color and mood

Name a palette and lighting direction. For brand work, compare the result with the approved color specification rather than judging by eye alone.

Common Mistakes to Avoid

10. Overcomplicating the prompt. Put the main subject and required copy first; add secondary details only if they matter.
11. Giving conflicting instructions. If you want a realistic 3D character, say so rather than asking for an undefined mix of photorealism and flat cartoon styling.
12. Using unsupported parameters. The WaveSpeedAI 4.5 text-to-image schema does not list the original draft’s negative_prompt, seed, quality, or n fields.
13. Ignoring aspect ratio. Choose the delivery shape before writing a prompt that depends on title placement or empty space.
14. Treating generated copy as approved copy. Verify names, prices, and claims against the source material.

Advanced Techniques

Layered prompt building: Assemble the subject, setting, lighting, and required text separately before combining them. This makes a failed result easier to diagnose.

subject = "A vintage bookshop"
setting = "floor-to-ceiling wooden shelves"
lighting = "warm ambient light"
text = 'a sign above the entrance reading "RARE BOOKS"'
prompt = f"{subject}, {setting}, {lighting}, {text}"
print(generate(prompt))

Style consistency across a series: Keep the style description fixed and change only the subject. Review whether the outputs actually belong together.

style = "minimalist digital illustration, flat colors, geometric shapes"
for subject in ["a coffee cup", "a laptop", "a desk plant"]:
    print(generate(f"{subject}, {style}"))

For the same product or character, move from prompt repetition to a reference-guided editing workflow.

Quality Optimization Checklist

Before generating:

• The prompt names the main subject and desired composition
• The output size suits the intended placement
• Required copy is quoted and assigned a location
• Reference images are included when exact details must survive
• The prompt separates what changes from what stays the same
• The review plan includes text, logos, proportions, and final-size inspection

FAQ

General Questions

Q: What makes Seedream 4.5 different from other AI image models?

A: ByteDance highlights its typography and reference-image editing improvements over 4.0. Whether that makes it the better choice depends on your prompt, input assets, and acceptance checks.

Q: Can I use Seedream 4.5 for commercial projects?

A: Review the current terms for the provider and your intended use. An API listing or a generated image alone does not establish the rights needed for every commercial project.

Q: How does pricing work?

A: As checked on October 30, 2026, WaveSpeedAI listed $0.04 per generated image for the 4.5 text-to-image endpoint. Check current pricing and billing for the exact endpoint and output count before a large run.

Q: Is Seedream 4.5 available elsewhere besides WaveSpeedAI?

A: Yes. ByteDance describes the model, and fal lists both a 4.5 text-to-image endpoint and a 4.5 editing endpoint. Use the documentation for the provider you select.

Technical Questions

Q: What image formats are supported?

A: WaveSpeedAI’s default output is a URL in data.outputs. Its schema also describes an optional base64 output mode. Inspect the returned file and current endpoint documentation before assuming a particular file format.

Q: Can I generate multiple images in one request?

A: The WaveSpeedAI sequential endpoint has a max_images parameter for a related set. The base text-to-image endpoint does not document the n field shown in the original draft.

Q: How do I ensure consistent results?

A: Keep the brief stable, use a reference-guided edit when the subject must remain the same, and compare outputs against a short list of required features. Prompt repetition alone is not a guarantee.

Q: What’s the maximum resolution?

A: It varies with the endpoint and current schema. 2048*2048 is WaveSpeedAI’s documented text-to-image default, not a universal maximum or a synonym for 4K.

Q: How long does image generation take?

A: Queue and inference time vary. The original 5–15-second estimate did not include a dated measurement method, so budget and measure against your own endpoint, size, and workload.

Q: Can I use custom aspect ratios?

A: The WaveSpeedAI text-to-image endpoint accepts a size string in width*height form, subject to its current limits. Start with a documented example if a custom size is rejected.

Text Rendering Questions

Q: Can Seedream 4.5 render text in any language?

A: ByteDance highlights dense text rendering, but that is not a guarantee across every language and script. Test the exact language, font treatment, and copy your project requires.

Q: Why doesn’t my text appear exactly as specified?

A: Image generation can alter characters or layout. Quote the copy, state its placement, generate alternatives, and check each candidate. Add final copy in a design editor when exactness is mandatory.

Q: Can I control font style?

A: You can guide the style with words such as “bold serif” or “handwritten.” For an exact brand font, apply the font after generation.

Q: How much text can be included in one image?

A: There is no reliable word-count rule in the cited model overview. Longer or smaller copy needs more careful checking. Test the actual layout rather than treating a short-text recommendation as a model limit.

Workflow Questions

Q: How can I integrate Seedream 4.5 into my application?

A: Submit to the chosen WaveSpeedAI model endpoint, save the prediction ID, poll its result endpoint, and read data.outputs when the status is completed.

Q: Can I batch process multiple images?

A: Yes. You can make separate requests programmatically, or use a sequential endpoint when you need a related set. Those are different workflows with different parameters.

Q: What if I’m not satisfied with the generated image?

A: Identify the failed requirement first: text, subject, layout, or style. Then change the relevant prompt instruction or use the edit endpoint with a reference image. Avoid changing every variable at once.

Q: Can I edit generated images?

A: Yes. Seedream 4.5 has a separate editing path. Supply accessible reference-image URLs and a prompt that states both the requested change and the details to preserve.

Q: How do I handle rate limits?

A: Retain the prediction ID and retry result polling after a temporary limit. If submission fails ambiguously, check for an existing task before submitting again.

Best Practices Questions

Q: What’s the best way to learn prompting for Seedream 4.5?

A: Start with one deliverable and a short acceptance checklist. Change one instruction at a time, and save the prompt beside the selected output.

Q: Should I use the highest resolution for all images?

A: No. Choose a size for the delivery format and increase it when the result needs more detail. Higher dimensions do not correct misspelled text or a changed product label.

Q: How can I maintain brand consistency?

A: Keep the visual brief and approved references together. For each output, check colors, product features, logos, copy, and composition against them.

Conclusion

Seedream 4.5 remains a useful version to evaluate for posters, text-led visuals, and reference-guided product edits. Its strongest documented claims come from ByteDance’s own 4.5 overview; a finished asset still needs a review against the brief.

Key Takeaways

Typography: Specify exact copy and proofread the output.

Editing: Use references and separate “change” instructions from “preserve” instructions.

API integration: Match each task to its endpoint and parameters; save the prediction ID and inspect the completed output.

Comparisons: Use the same prompt and acceptance criteria across models. Do not rely on an undated rank or a blanket “best” claim.

Getting Started

1. Choose one real deliverable: a poster, product edit, or related image set.
2. Select the matching Seedream 4.5 endpoint.
3. Start with a documented size and a focused prompt.
4. Generate a candidate and inspect it at delivery size.
5. Save the prompt, references, endpoint, and result.
6. Compare the accepted result and correction effort with your other options.

Future Outlook

Newer Seedream versions are available as of October 2026. Keep a 4.5 workflow when it meets your quality, cost, and compatibility needs. Consider a newer version when 4.5 repeatedly misses a requirement—such as preserving a label or handling a dense layout—and compare both versions on the same inputs before migrating. Existing 4.5 endpoints do not become 5.x endpoints simply because a newer model exists.

---

Previous posts:

• FLUX 3 Max guide
• Comfy Cloud review
• Comfy Cloud alternatives
• ModelsLab review
• Holo4 API vs local deployment