Day 2: The Image/Video Model Landscape, Cinematic Prompting, Cost Strategy, and a Commercial Built Live

Year by year: 2022 was the ChatGPT moment and the 'AI copywriter' era; 2023 Midjourney (born in Discord) made images mainstream while Runway was an experimental video model; 2024 video was still 'very experimental'; 2025 tools became 'ready for prime time' for commercial use and workflows emerged around them. 'It's still very early.' The landscape at recording - images: Midjourney, Google Nano Banana / Pro, ByteDance Seedream, OpenAI ChatGPT Image 1.5, Black Forest Labs Flux 2 (open), Ideogram (text rendering), Recraft, Qwen; video: Runway, Kling, Google Veo 3.1, Grok, Sora 2 (social, auto multi-shot, less controllable), Moon Valley (licensed data only), LTX Studio, WAN 2.2 Animate (full-body character replacement), Hailuo.

For a neutral ranking he uses Artificial Analysis: two anonymous images, people vote left or right, no brand bias - the day before, ChatGPT Image 1.5 had edged Nano Banana Pro. His own kit: Seedream, Nano Banana Pro, Krea's model, Flux for stills; Seedance, Kling ('love Cling'), Runway, Hailuo for motion.

The capability cluster that arrived late 2025, credited to Google's Nano Banana and Nano Banana Pro ('about a month or two ago'): natural-language image editing, single- and multi-image references, changing camera angle on an existing image, start/end-frame video generation, and video with synchronized spoken dialogue. Multi-image reference - 'replace the man in the first image with the man in the second image' - is what lets you put a real person or a real product into a generated scene.

This is what makes 'character consistency, scene consistency, product consistency' achievable, which he calls 'a massive innovation in the last 6 months to the past year.' The working technique: never prompt the next shot from scratch; feed the previous generation in as the reference so wardrobe, face, and set carry through the sequence.

Image prompts: Subject -> Action -> Scene/Environment -> Atmosphere -> then camera language (camera body, lens, angle, lighting, film stock, style). 'You're not writing full sentences... separate every single element by commas,' and front-load what matters because 'whatever you have near the front is going to have the most emphasis.' Leave lighting or style out and 'it will just guess for you.'

Cheat sheet: Dutch angle (tilted, unease), low angle (power), close-up / medium / cowboy (thigh-up) / extreme close-up; lighting - natural, volumetric, golden hour, rim, bioluminescent; lens - higher mm is tighter (50mm portrait vs 10-20mm wide); film stock for a retro look; real camera names work (Sony A7S III, ARRI Alexa, RED, Canon 1DX). Video prompts are different - 'just describe the scene that you wanna see' in sentences. Edit prompts are imperative: 'show me, remove, replace, place him a meter back.'

Realistic casual spend is '20 to 50 dollars per month' in total. Two aggregator models: monthly subscription with a nice UI and expiring credits (Freepik, Krea), or pay-per-use with a rawer interface (fal.ai, Replicate) - cheaper overall. Free trials are for 'learning and exploring... not good for producing a real piece of content': limited credits, watermarks on video, queues - though you can stack trials (Hailuo, Kling, Higgsfield) and scale a video to 110-120% to crop a corner watermark.

Open vs closed via an Android/Mac analogy: open models (Flux, ComfyUI locally) can be fine-tuned with a LoRA - training on brand colours, product photos, or a person's likeness for consistency; closed models cannot. Midjourney is the odd one out - no public API, so it must be paid for directly. Free tools list: Google ImageFX, ElevenLabs (voice + music), CapCut, DaVinci Resolve, Canva. ComfyUI exists but 'too hard for me to look at'; Flora and Weavy are friendlier node-based tools.

Three phases named and used: pre-production (storyboard), production (generate images, then video, then music/voice), post-production (edit). The story: 6 AM alarm, zombie shuffle to the kitchen, brew 'Billin' Coffee - Ultra Strong, Zombie Wake Up', sip, lightning transformation into a suited professional, line: 'That's the best coffee I've ever had.'

What the demo teaches beyond the steps: test the same prompt across models when one keeps failing (the alarm clock rendered wrong until Seedream 4.5); insert yourself by multi-image reference and describe wardrobe explicitly or the model copies the reference's clothes; generate at 1K while iterating and upscale to 4K only for the keeper; chain outputs as references; generate coverage (extreme close-up of the portafilter, overhead) so the edit has options; Kling 2.5 Turbo for speed, 15 credits per 5s clip; Flow/Veo 3.1 for the one shot that needs spoken dialogue and timestamped action; and the storyboard is a reference only - it cannot be imported into any generator.

OpenAI drew backlash for lax guardrails (full South-Park-style episodes), then blocked 'famous people' outright; Disney then invested in OpenAI to license its characters for legal generation. Meanwhile Disney is suing Midjourney (mid-2025), joined by Warner and Universal, and another suit targets a company heard as 'Haylou' (possibly not the video tool). The pattern he expects, borrowed from music (Sony, Universal, Warner vs the music generators): sue, then settle into partnership and investment - 'a big theme of 2026, that the dust is gonna start settling.'

Practical consequences from Q&A: all tools accept your own photo/video but filter well-known faces; starting from your own drawing does not create stronger legal ownership - 'AI law is unsettled' - it just changes the creative input.