Buttmunch away beccazican, its interesting discussion. But also share more of your stuff because its awesome and its an art topic. 
Hey, no worries. I really appreciate the discussion.
I don’t agree with the “editor” view, though. That perspective makes sense if your experience is limited to ChatGPT, but not once you’ve worked with dedicated image models and the thousands of available plugins, scripts, and tools. The level of control and creative influence goes far beyond simple editing.
With that much hands-on control, would be like dismissing photographers or film directors as as “button pushing editors” instead of artists.
This whole field is still new, and most people only see it through surface-level media coverage. It’s similar to the 1800s, when some folks considered photography a soulless mechanical trick and a perversion of art, yet today no one doubts it’s an art form.
The hardest thing to grasp is the depth beneath the surface. Both AI workflows and standard creative software are moving toward conversational tools. Soon you’ll tell Photoshop what you want instead of clicking every button, but that won’t make it less powerful or valuable, the complexity still runs underneath.
And most importantly, image models don’t copy their training data. They learn patterns and combine them to create something new. It’s like looking at a few pictures of Pikachu online, then drawing Pikachu in an astronaut suit on the moon from memory: You’re using what you’ve learned, not copy-pasting it.
But creation alone doesn’t make something art. A plain portrait isn’t art by itself either; it’s the style and idea, whether creative, funny, ironic, critical or thought-provoking, that turns technique into artistry.
PS: It’s an interesting discussion, but I think we should continue it in the AI Art Megathread since it’s probably too off-topic here.
4 Likes
I’ve created a separated thread and moved all relevant comments to it, as requested.
Everyone please use:
- This thread for AI artworks
- The existing E4 Pokemon fan art and crafts thread for all other crafts/arts
- The existing The Great AI Art Debate for ethical AI-art discussions
Greetz,
Quuador
13 Likes
Is the amount of freshwater and electricity and processing power used for non-artists to label themselves artists worth it? I think not but the wastefulness is the hallmark of our current society.
2 Likes
You know, that’s a really important question. There’s a very widespread narrative that AI is destroying the planet, so I actually spent some time digging into the data on the sustainability of AI.
Let me summarise this for you:
Peer-reviewed data already show that AI creation is far less resource-intensive than human creation.
A Nature study (one of the most prestigious scientific journals in the world) found that generating a written document with AI emits 130–1500 × less CO₂e per page than a human writer, and that AI illustrations emit 310–2900 × less CO₂e per finished image.
Energy + water per query: According to OpenAI CEO Sam Altman’s June 2025 disclosure and Epoch AI’s independent analysis, a typical ChatGPT / GPT-4o-style query uses about ≈0.30–0.34 Wh of electricity and 0.000085 gal ≈ 0.32 ml of water per query. Epoch also shows that the widely cited 3 Wh/query figure is likely an overestimate for modern deployments.
Smaller models use far less: Many popular text-to-image and diffusion models have much smaller active parameter counts than frontier LLMs and run on single GPUs. Empirical measurements and recent analyses place typical image generations in roughly the ~0.01–0.1 Wh per image range for efficient open source models, corresponding to costs of 3.6–36 seconds of a 10 W LED bulb per image (Source 1, Source 2).
Compare that to traditional workflows:
• Producing one A4 sheet of paper consumes roughly 2–13 liters of water, equal to about 6,000–40,000 ChatGPT queries at 0.32 ml each.
• Every art material carries a footprint: paints, pigments, binders, solvents, canvases, paper, brushes, and packaging all require raw materials, chemicals, and energy-intensive manufacturing.
• Traditional pigments and mediums such as cadmium, cobalt, or solvent-based oils can introduce heavy-metal toxicity and persistent residues if not properly handled or disposed of.
• Paper and canvas production involves forestry or cotton cultivation, pulping, bleaching, coating, and global transport.
• Studios and galleries use electricity for lighting, heating/cooling and ventilation
• “purely digital” traditional workflows (tablet + Photoshop) incur electricity consumption multiplied by long working hours (also see figures above regarding carbon footprint)
Training costs for frontier models are indeed large: Ren et al. estimate that training GPT-3 could evaporate around 700,000 liters of freshwater in U.S. data centers. Based on a water footprint of roughly 2–13 L per A4 sheet, this corresponds to about 50,000–350,000 sheets of A4.
For comparison, using an annual print run of ~125 million copies, the New York Times prints on the order of 6.85–13.7 million sheets per day (assuming an average of 20-40 sheets of paper / copy), so GPT-3’s entire training run is roughly 0.4–5% of a single day’s NYT paper output on a sheet-for-sheet basis. And this is a conservative overestimate, because a broadsheet newspaper page is larger and heavier than A4 and thus requires more pulp, water, and energy.
Put differently: for the paper the NYT consumes in a single day, you could train at least 20 to 250 ChatGPT-scale models a day.
Crucially, that training cost is a one-time hit amortized over billions of inferences. Moreover, dedicated image models are one to two orders of magnitude smaller than GPT-3, making their training runs vastly less resource-intensive.
The conclusion: If you’re genuinely concerned about sustainability, or, as you put it, about the “wastefulness of our current society,” then by your own logic you’d have to argue for AI art, not against it. (I’m not doing that here; I’m simply following your reasoning to its logical end.)
Super Mew vs Joker Mewtwo
Image Generation Workflow – Super Mew vs. Joker Mewtwo
-
Base Generation: Created locally in ComfyUI using QwenImage with reference images of Mew and Mewtwo to define their shapes and appearance. Generated the first version of the scene via prompt instructing the model to generate Mew in a Superman costume fighting Mewtwo as the Joker. Corrected visual errors with QwenImage Edit 2509 and Photoshop. (~1100 × 800 px)
-
First Upscale: Applied the Flux Dev model with Nomos 8K ADT JPG (×1.5), adding ~0.35 noise to enhance resolution and fine details.
-
Second Upscale: Used Stable Diffusion 1.5 (CyberRealistic v3.1 + Detailer LoRA) for additional refinement and another ×1.5 upscale, optimized with HyperTile for efficient high-resolution rendering.
-
Final Enhancement: Performed a 2× upscale in Gigapixel Topaz, reaching a final resolution of ≈ 5000 × 3700 px before cropping.
1 Like
Why is it not batman mew. AI should know better.
1 Like
1 Like