AI image generation for arbitrage creatives: Midjourney, DALL-E, Flux - practical guide

AI image generation in 2026 is not a toy for designers or entertainment on Discord. For an affiliate marketer, this is a full-fledged conveyor of visuals: Midjourney v6, DALL-E 3, Flux Pro, Stable Diffusion XL and Leonardo AI produce creatives in minutes that previously took days and thousands of dollars to shoot. Static banners, references for video generation, elements for slide shows and motion-creation - AI images have become the foundation of production in all verticals, from nutra to gambling. But this approach has pitfalls: AI detection on platforms, stylistic consistency between campaign creatives, and most importantly, scaling. When AI images become the basis of video creatives for a network of accounts, without uniqueness, scaling becomes a lottery. In this article, we analyze each tool, build prompts for specific verticals, and show the full path from a text request to a unique video creative ready for upload.

Tool overview: Midjourney v6, DALL-E 3, Flux Pro, SDXL, Leonardo AI

The AI image generation market has changed radically over the past year and a half. If in 2024 Midjourney dominated with almost no competitors, then in 2026 an affiliate marketer can choose from five tools, each with its own strengths. Let's look at them from the point of view of practical value for creating creatives.

Midjourney v6

Market leader in quality and aesthetics. Midjourney v6 produces images that are literally indistinguishable from studio photography - the right light, natural skin texture, cinematic composition. For nutra-creatives and dating, this is the gold standard.

Access: works through a Discord bot or through its own web interface (midjourney.com). The web version appeared in 2025 and greatly simplified the workflow - no more messing around with Discord channels. The API is available for commercial users.

Pricing: Basic - $10/month (~200 generations), Standard - $30/month (~900 generations), Pro - $60/month (unlimited in relaxed mode, 30 hours fast). For arbitrage volume, Standard or Pro is optimal. The cost of one image in fast mode is about $0.03–0.07.

Strengths for arbitrage: photorealism of faces and bodies (critical for interiors and dating), custom styles via the --style and --sref parameters (campaign consistency), upscale up to 4K, variations of one image via the --v parameter. The --sref (style reference) parameter allows you to set a reference image, and all subsequent generations will follow its visual style - invaluable for a series of creatives in one campaign.

Limitations: strict content moderation - rejects prompts with medical claims, explicit content and some gambling topics. Does not generate text reliably (letters are distorted). No direct API for mass automation without commercial subscription.

DALL-E 3 (OpenAI)

The main advantage of DALL-E 3 is the accuracy of following the prompt. While Midjourney often “interprets” a request in its own way, adding beauty at the expense of precision, DALL-E 3 does exactly what you ask. For an affiliate marketer who needs a specific scenario in a frame, this is critically important.

Access: via ChatGPT Plus/Pro, via OpenAI API, built into Microsoft Designer and Bing Image Creator. API access is the most flexible option for mass generation: can be automated through scripts.

Pricing: via ChatGPT Plus ($20/month) - limited number of generations. Via API - $0.04 for a 1024×1024 image (standard quality) or $0.08 for HD quality. For 100 creatives via API - $4–8. The most predictable and transparent pricing on the market.

Strengths for arbitrage: best on the market work with text on images (inscriptions, labels, call-to-action - generated in a readable manner), precise adherence to compositional instructions, native integration with ChatGPT for iterative refinement of prompts, the ability to edit individual areas of the image (inpainting).

Limitations: photorealism is inferior to Midjourney v6 - images look a little more “digital”. Strict moderation by OpenAI - rejects public figures, medical content, and gambling. Built-in C2PA watermark in metadata (easily removed, but you need to remember). Maximum resolution - 1024x1792 (no upscale to 4K native).

Flux Pro (Black Forest Labs)

Flux Pro is a dark horse of the market, which in 2025 has gained a critical mass of users among arbitrage traders. The reason is simple: excellent quality with minimal censorship and an affordable price. This is an open-source architecture with a commercial API - a combination that gives maximum flexibility.

Access: Flux Pro via API (fal.ai, Replicate, Together AI and other hosting), Flux Dev and Flux Schnell are free models for running locally. Local launch of Flux Dev on a video card with 12+ GB VRAM - completely free generation without limits and censorship.

Pricing: Flux Pro via API - $0.055 per image. Flux Schnell (fast version) - $0.003 per image. Local launch Flux Dev - free (electricity only). For mass testing of hypotheses through Flux Schnell, 1000 images will cost $3.

Strengths for arbitrage: minimal built-in censorship (especially in local versions - it generates almost everything), quality at the level of Midjourney v6 in the latest versions, support for LoRA adapters for training on your data (you can “teach” the model to generate a specific product or style), lowest cost on the market. For the gambling vertical and aggressive nutra-creatives, this is the best choice precisely because of the lack of strict moderation.

Limitations: there is no convenient web interface of the Midjourney level (API or local launch requires technical skills), local launch requires a powerful video card, text on images is generated worse than that of DALL-E 3.

Stable Diffusion XL (Stability AI)

SDXL is a workhorse for those who want total control and zero running costs. A completely open-source model that runs locally and generates without any restrictions. SDXL ecosystem - thousands of custom models, LoRA adapters and extensions on CivitAI.

Access: local launch via ComfyUI, Automatic1111 or Forge. Cloud - through API providers (Stability AI API, Replicate). For full operation, you need a video card with 8+ GB VRAM (optimally 12–16 GB).

Pricing: local - free. Via Stability AI API - $0.03–0.06 per image. The only investment is time to set up the environment (ComfyUI) and a powerful video card if you work locally.

Strengths for arbitrage: absolute freedom of content (no censorship in local mode), a huge library of custom models on CivitAI (there are specialized models for beauty, lifestyle, product photography), ControlNet for precise control of pose and composition, batch generation via ComfyUI workflow - you can generate hundreds of options on autopilot.

Limitations: basic SDXL quality is inferior to Midjourney v6 and Flux Pro (but custom checkpoints close the gap), requires technical knowledge to configure, slower than cloud services when generating on consumer video cards.

Leonardo AI

Leonardo AI is the most accessible entry point for beginners. A convenient web interface, a generous free plan and a set of ready-made models tailored to specific styles - from photorealism to animation.

Access: leonardo.ai web interface, API for paid subscribers. Registration is free, without linking a card.

Pricing: free plan - 150 tokens/day (enough for 30-50 images). Apprentice – $12/month (8,500 tokens). Artisan – $30/month (25,000 tokens). Maestro – $60/month (60,000 tokens). The free plan is often enough to test hypotheses.

Strengths for arbitrage: pre-trained models for specific styles (PhotoReal, DreamShaper, Anime), built-in editor for inpainting and outpainting, AI Canvas for combining several generations, generation of textures and UI elements - useful for gambling-creo. Generous free testing plan.

Limitations: photorealism is noticeably inferior to Midjourney v6 and Flux Pro, limited control over style compared to SDXL, content moderation (softer than DALL-E, but stricter than Flux).

Prompt engineering for arbitrage creatives: formulas for verticals

The prompt is 80% of the result. An affiliate marketer who has mastered prompt engineering receives a pipeline of visuals without a designer. Below are proven formulas and approaches for key verticals. Each formula is tested on real campaigns and adapted to the specifics of AI generators in 2026.

Universal structure of prompt

Regardless of the vertical, an effective prompt for an AI image is built according to the formula:

• Subject: who or what is in the frame (person, product, scene)
• Action/posture: what the subject does
• Environment: where the scene takes place
• Shooting style: camera type, angle, depth of field
• Lighting: type and direction of light
• Mood/atmosphere: emotional tone of the image
• Technical parameters: resolution, aspect ratio, style modifiers

Each element of the formula adds control over the result. Skip elements deliberately - AI will fill in the gaps at its discretion, and the result will be less predictable.

Nutra: before/after, food photos, lifestyle

Nutra requires maximum photorealism. The viewer must believe that they are seeing the real result of a real product. The best tools are Midjourney v6 and Flux Pro.

• Before/after: generate two separate scenes rather than one split image. Prompt for “before”: ask a realistic problem (acne, dull skin, excess weight) without being cartoonish. Prompt for “after”: same location, similar angle, but improved version. Use --seed in Midjourney or fixed seed in Flux for maximum similarity between two images
• Product photo: “Professional product photography of [product description] on a marble surface, soft studio lighting, shallow depth of field, clean white background, commercial photography style, 8K” - this template works for any nutra product
• Lifestyle shots: “Young woman applying face cream in a bright bathroom, natural morning light through window, close-up shot, Canon EOS R5 85mm lens, beauty editorial style, soft bokeh” - camera and lens detail helps the AI generator reproduce a realistic photography style

Critical: Do not use medical claims in prompts - Midjourney and DALL-E will block them. Instead of “anti-aging cream that removes wrinkles” write “woman with radiant glowing skin, luxury skincare product” - the result is the same, censorship does not work.

Dating: lifestyle, emotions, attractiveness

Dating creatives are about emotions and atmosphere. AI generation works great here because the neural networks are trained on millions of lifestyle photos.

• Lifestyle portraits: “Attractive [man/woman] in casual outfit at a rooftop café during golden hour, candid smile, warm natural lighting, shallow depth of field, iPhone photo style” - iPhone photo style increases confidence because it looks like a real selfie or photo of a friend
• Emotional scenes: “Couple walking on the beach at sunset, holding hands, back view, romantic atmosphere, warm color palette, cinematic wide shot” - generate scenes that make you want the same life
• For Kling AI lip-sync: generate a portrait using Midjourney or Flux, then use it as a reference in Kling AI to create a “talking” video - CTR such bundles are 2-3 times higher than static banners

Recommendation: for dating, generate a variety of types - different ethnicities, ages, clothing styles. This expands the audience pool and allows you to A/B test which type resonates with a specific GEO.

Gambling: luxury, dynamics, neon

Gambling creatives are the exact opposite of visceral: realism is not needed here, but visual shock and a feeling of luxury are needed. The best tools are Flux Pro (minimal censorship) and Leonardo AI (ready-made style presets).

• Luxury lifestyle: “Luxury penthouse interior at night, neon purple lighting, gold accents, expensive whiskey on a glass table, casino chips scattered, cinematic dramatic lighting, wide angle shot” - motivating context without direct image of gambling interfaces
• Winning emotions: “Person celebrating with arms raised, gold confetti falling, neon lights background, euphoric expression, dramatic low-angle shot, nightclub atmosphere” - the emotion of winning is more important than the casino image itself
• Abstract elements: “Golden slot machine symbols floating in dark space, neon glow effects, 3D render style, luxury casino aesthetic, dramatic volumetric lighting” - for overlay elements in video creatives

For gambling, Flux Pro is the preferred tool because Midjourney and DALL-E often reject prompts mentioning casinos, betting, and gambling. Flux Pro (especially in the local version) generates without restrictions.

Style consistency: how to maintain a single campaign visual

One of the main problems with AI generation for arbitrage is that each image looks like a separate work. But effective creative requires stylistic unity: all visuals of one campaign should be perceived as part of one series. The viewer sees the ad three to five times before conversion—and must recognize the brand each time.

Consistency tools

• Midjourney --sref: style reference parameter - you pass the URL of the reference image, and all subsequent generations will inherit its color palette, lighting and general aesthetics. Works as a “visual brand book” - set up once, hundreds of generations in the same style
• Midjourney --cref: character reference - for generating the same “character” in different situations. Not perfect, but significantly improves facial consistency from frame to frame
• SDXL + LoRA: train the LoRA adapter on 10–20 images of the desired style - the model will generate in this style endlessly. The training process takes 30–60 minutes on an RTX 3070+ level video card. For advanced arbitrage traders - the best style control tool
• Flux + IP-Adapter: analogue of --sref for Flux - allows you to set a style reference through an additional image. Works through ComfyUI nodes
• Leonardo AI Styles: pre-trained style presets - select once, apply to all campaign generations. The simplest option for beginners

Practical work flow with style consistency

Recommended approach for an arbitrage campaign:

• Step 1: Generate 20-30 images with different style settings. Select 3-5 that look like series
• Step 2: Use the best image as --sref reference (Midjourney) or as input for IP-Adapter (Flux/SDXL). All subsequent generations will be in this style
• Step 3: Fix the seed value for elements that should be repeated (color scheme, lighting, background)
• Step 4: Post-processing - a single LUT (color lookup table) on top of all images in the series for final color harmonization. Free LUTs for Photoshop or DaVinci Resolve will cover this task

Style consistency is also important when creating video creatives from AI images. When a series of images is turned into a slideshow or image-to-video animation, style jumps between frames reduce perception and kill conversion. Keep a single visual and your creatives will look professional.

AI image detection: risks and workaround strategies

In 2026, all major platforms have implemented AI content detection systems. This affects not only video (which we wrote about in detail in the article about AI video generation for affiliate marketing), but also static images. For an affiliate marketer who uses AI visuals in creatives, understanding detection mechanisms is not a theory, but a necessity.

How platforms define AI images

• Metadata C2PA / Content Credentials: DALL-E 3 and Adobe Firefly embed digital watermarks in file metadata. Platforms read these tokens automatically. Deletion - by resaving the image without EXIF or converting the format (PNG → JPG → PNG)
• Neural network classifiers: Meta uses models like SSCD, Google uses SynthID. They analyze patterns at the pixel level: the characteristic “smoothness” of textures, the specific structure of noise, the unnatural regularity of micro-details. Each AI generator leaves its own “spectral fingerprint” - and classifiers are trained to recognize it
• Statistical analysis: AI images have a different frequency distribution in the Fourier spectrum compared to real photos. This is a thin but reliable marker that is difficult to remove with conventional processing

Strategies for reducing AI detection

It is impossible to completely eliminate detection - algorithms are improving faster than bypass methods. But it is possible to significantly reduce the probability:

• Post-processing in Photoshop / Lightroom: adding grain, chromatic aberration, slight motion blur. These artifacts are typical of real photographs and confuse AI classifiers. Don't overdo it - the image should look natural and not spoiled
• Combining AI and real elements: use an AI background + a real product photo, or a real person photo + an AI-generated environment. Hybrid images are much more difficult for classifiers
• Resize and Recompress: Reduce the image to 70-80% of the original size, then resize it back. This introduces compression artifacts characteristic of real photos sent through instant messengers and social networks
• Metadata cleanup: complete removal of EXIF, XMP and IPTC data. Regenerating a file via the canvas API or resaving with random compression parameters
• Using SDXL / Flux instead of DALL-E: open-source models do not embed C2PA markers. One detection level is removed immediately

It is important to understand: for an affiliate marketer, AI image detection is half the problem. The second half is when AI images become part of the video creative (slideshow, animation, image-to-video) and this video creative is uploaded to the account grid. Here, double detection is added to AI detection: the same video on 20+ accounts connects the entire network, and the same AI artifacts based on the video strengthen the signal for antifraud.

Workflow: from AI image to unique video creative on the platform

AI image is a raw material, not a final product. In traffic arbitrage, images are almost always turned into video content: slideshows, animated banners, image-to-video, motion collages. Here is the complete pipeline from a text prompt to a unique video creative, ready to be uploaded to the platform.

Stage 1: Mass Generation and Selection

Start by defining the visual strategy of the campaign: vertical, GEO, target audience, type of creative. Study competitors through spy services - which visual solutions are working right now.

Generate with reserve: 50–100 images for 10–15 prompts. Use 2-3 tools in parallel (for example, Midjourney + Flux Pro + Leonardo AI) - different models give different results, and this expands the pool of quality visuals. Budget for generating 100 images – $10–30.

From 100 generated images, select the best 15–25 according to the criteria: photorealism, absence of artifacts (check hands, textures, background), compliance with the creative script, emotional strength.

Stage 2: Image post-processing

Each selected image is processed:

• Removing small artifacts through inpainting (DALL-E 3, Leonardo AI or Photoshop Generative Fill)
• Overlaying text elements: headings, CTAs, price tags - AI generators do not render text well, add it manually in Photoshop or Canva
• Color correction to a single campaign style (single LUT for the entire series)
• Adding grain, chromatic aberration and other “realistic” artifacts to reduce AI detection
• Metadata clearing from C2PA markers and AI signatures

Stage 3: Transformation into video creative

Here AI images are transformed into video content - the main format for TikTok, Reels and Shorts:

• Slideshow with effects: series of AI images with Ken Burns effect (zoom + pan), transitions and text overlays - CapCut, DaVinci Resolve or even built-in tools TikTok
• Image-to-video: AI image as a starting frame for AI video generation via Kling AI, Runway Gen-3 or Sora 2. The result is a 5-15 second video where a static image “comes to life”
• Motion collage: several AI images + animation + text + audio - in the format of a dynamic advertising video
• Adding audio: trending sound, AI voice acting (ElevenLabs, LOVO) or background music - critical for coverage on all short-video platforms

Stage 4: Uniqueness for the account grid

This is the critical point that separates the amateur from the professional. You have 10–15 ready-made video creatives, and you need to upload them to 30–50 accounts. Each account must receive a technically unique version - otherwise the content bundle will kill the entire network.

360° Uniquizer takes each video creative and creates N unique versions from it - as many as there are accounts in your network. Each version is unique at all verification levels:

• Perceptual hashes: color space shift, geometric transformations, crop - the hash of each version is different from the original and from all other versions
• Audio fingerprint: transformation of the audio track - pitch, tempo, background noise - each version sounds different for algorithms, but the same for humans
• Neural network analysis: restructuring editing, inserting micro-frames, changing scene timing - the platform’s AI detector sees different content
• AI patterns: pixel transformations disrupt characteristic artifacts of neural network generation - the likelihood of AI detection is reduced with each unique version
• Metadata: complete regeneration - bitrate, codec, timestamps. Each version is technically a new file

Output: from one AI video creative - 20, 50, up to 200 unique versions. Each is verified as original content. Double protection: against double detection and against AI detection at the same time.

Phase 5: Distribution and Analytics

Unique versions are distributed across accounts - each account receives its own unique file. Flood at natural intervals, at different times of day, with different order of content - a complete imitation of organic activity.

After 24–48 hours, collect analytics. You scale creatives with the best metrics (CTR, conversion, reach) - generate variations of the same prompts, create new video versions and uniquize them to expand the network. Ineffective ones - replace them. The cycle “generation → processing → video → uniqueness → upload → analytics” is repeated continuously. AI generation makes each iteration of the cycle lightning fast: a new visual in a minute, a new video creative in 10 minutes, uniqueness of the package in another minute. The full cycle of testing a new hypothesis is one working day, not a week, as with manual production.

This workflow works for all short-video platforms: TikTok, Instagram Reels, YouTube Shorts, Pinterest Video. The difference is in the formats and aspect ratios, but uniqueness through 360° Uniquizer is equally effective for any format.

Economics: AI images + video + uniqueness vs traditional production

Specific numbers for a typical affiliate marketer’s task: you need 50 unique video creatives of 15 seconds each for a grid of 30 accounts in the nutra vertical.

Traditional approach

• Photo shoot with model and product: $500–1,500
• Video shooting: $500–2,000
• Installation of 50 rollers: $750–1,500
• Manual diversification for 30 accounts: 3–5 hours of manual labor, incomplete uniqueness
• Time: 1–3 weeks
• Total: $2,000–5,000+ and weeks of waiting

AI-pipeline

• Generate 100 images (Midjourney + Flux Pro): $15–40
• Post-processing: 2–3 hours of your own time
• Image-to-video via Kling AI / Runway: $20–60
• Editing video creatives: 3–4 hours
• Unique via 360° Uniquizer — 10 videos × 5 versions = 50 unique files: processing minutes, software license cost
• Time: 1–2 days
• Total: $50–150 + license 360° Uniquizer and 1–2 days of work

The difference is 15–30 times in cost and 5–10 times in time. And most importantly, with the AI ​​approach, the iteration cycle is compressed to hours. If creative is burned out, a new one is ready on the same day, not in a week. This radically changes the economics of arbitrage: instead of one expensive bet on the “right” creative, you test dozens of hypotheses in parallel and scale only what works.

Optimal strategy: budget tools (Flux Schnell, Leonardo AI free plan) for mass testing of prompts → regenerate the best prompts into Midjourney v6 / Flux Pro for maximum quality → turn the final creatives into videos → uniquize them through 360° Uniquizer for the entire grid. Minimum costs at the testing stage, maximum quality at the scaling stage.