Six AI design tools tested, editable layers are the real divide

Last month we ran a cost comparison of image models (batch 22) and an AI video workbench comparison (batch 33). This time the angle changes: in a design workflow, one good-looking image is only the start; whether you can keep editing it, split it into layers, and self-host it decides if it enters production. Six tools are in the ring: Ming-Image-0.1-Design+Layer, Midjourney, Jimeng AI, Canva, Qwen-Image 2.1, and Nano Banana 2 (Gemini 3.1 Flash Image). Settle the accounting first.

What This Comparison Covers, and What It Deliberately Skips

Four rules before any tool is judged.

First, cost and license shapes cannot be compared head to head. Among the six we have an MIT self-hosted model, a Qwen non-commercial model, closed SaaS, and a per-image API. Their denominators differ: local is GPU depreciation plus power, subscription is monthly, API is per call. Every cost item states its basis; we crown no "cheapest".

Second, several benchmarks are vendor or review sourced, not independently retested. Nano Banana 2 pricing, Ming-Image's "about 20 seconds" and "edit cost 1/7", and the Crello "12 of 12, 4.3x" all come from ai-bot and orcarouter. This site did not reproduce them, so each is tagged accordingly.

Third, the division with batch 22 cost comparison: that piece is price and unit quality; this piece is workflow, editability, layers, RGBA, licensing, and self-hosting, not pure quality.

Fourth, the division with batch 33 video workbench: that piece is video and real-time; this piece is static design images, not video.

The Six Tools at a Glance

Ming-Image-0.1-Design+Layer (Ant Group inclusionAI, MIT). The only open-source design line here. Two models: Design does text-to-design for UI, dashboards, infographics, and posters end to end; Design-Layer decomposes a flattened design into 2 to 9 independent RGBA layers you can change, move, or swap. Repo inclusionAI/Ming-Image on GitHub, weights on HuggingFace and ModelScope. MIT is confirmed across four sources (GitHub API, datanorth.ai, bittide, orcarouter) with commercial use allowed.

Midjourney (closed-source SaaS). Strong aesthetics and mature stylization, the benchmark for concept art, but it returns a flat image with no native layers, and text rendering often needs fixing.

Jimeng AI (ByteDance, closed-source). Chinese-friendly and fast, suited to short-video covers and social artwork. Layout is moderate, no editable layers, consumer-leaning.

Canva (design platform, templated). Not a generation model but a platform: templates plus AI assistance, mature layer editing, but creativity comes from templates, not end-to-end generation.

Qwen-Image 2.1 (Alibaba, Qwen Research License, non-commercial). Roughly 7.1B single-stream DiT with native RGBA and stable Chinese-English layout. The catch is licensing: the Qwen Research License is non-commercial, a hard line against Ming-Image's MIT.

Nano Banana 2 = Google Gemini 3.1 Flash Image (closed-source API). Priced per image, output up to 4K with SynthID and web retrieval. Text and layout are stable, but it returns one image with no layers. Pricing: about $0.067 per 1024-squared, $0.101 per 2048-squared, $0.151 per 4K, batch roughly half, text input $0.25 per million tokens (orcarouter vendor basis, not independently retested).

The Head-to-Head Table

This table is the spine of the piece. All five dimensions center on "design workflow / editability", not pure image quality.

Reading the table: only Ming-Image and Qwen-Image 2.1 provide native RGBA, but their licenses point opposite ways; Midjourney, Jimeng, and Nano Banana all return flat single images; Canva's "editability" comes from templates, not a generation model. In the cost column, local compute, subscription, and per-image billing are three shapes that cannot be directly summed, so this piece crowns no "cheapest". For a team shipping real assets, that contrast is the whole ballgame: a flat image locks every later change behind a full regeneration, whereas a layer file keeps the door open, and we build on that in the next section.

Text Rendering and Layer Editability

A design image differs from "one pretty picture" in that it usually carries text, layout, and hierarchy. This section is the core of the whole piece.

Text rendering. Posters, infographics, and UI fear AI garbling words or breaking the grid. Ming-Image organizes requests into copy, module, layout, and style dimensions, supports 8K structured prompts, and absorbs long input intact; from the card and README, text and layout are stable (model card basis, not retested). Nano Banana 2 also renders text and layout stably and retrieves from the web. Qwen-Image 2.1 holds Chinese-English layout steady. Midjourney's aesthetics are strong but text stays the weak spot, often needing later fixes. Jimeng is Chinese-friendly but weak on complex layouts. Canva does not generate, so text is controllable, at the cost of template-bound creativity.

Layer editability is the real dividing line. After a design ships, a client says move the title, swap the background, export this icon. Ming-Image's Design-Layer uses Type Tokens for each layer's role, Alpha-Aware Layer Optimization for transparent edges, and Composite-Layer Stack Consistency to reconstruct the original, yielding 2 to 9 RGBA layers you operate separately. With the PPT Skill, a design restores to editable PowerPoint in one step, a chain no competitor offers. Midjourney, Jimeng, and Nano Banana return flat images, so one word change often means full regeneration; Canva's layers live inside its template system, not generated semantics; Qwen-Image 2.1 has RGBA but no public Layer-style decomposition (per official). If editing is frequent, native RGBA plus auto layering is irreplaceable; for an atmospheric image, flat output suffices.

Self-Hosting, Licensing, and Cost Shape

This section clarifies "can you control it yourself" and "how much does it cost, and in what name".

Self-hosting and the license red line. Ming-Image is the only one with clear MIT, self-hostable and commercial. The GitHub API snapshot (2026-09-24) shows 91 stars, 6 forks, Python, MIT, with four sources confirming commercial permission. Qwen-Image 2.1 runs locally but the Qwen Research License is non-commercial. Midjourney, Jimeng, Canva, and Nano Banana are closed-source, control on the vendor side. The gap matters: two models output RGBA, yet one sells and the other only researches.

Hardware threshold (model card). Minimum validated deployment is a single GPU with at least 80 GiB VRAM at BF16; 24 GiB consumer cards have no official guidance, community quantization (INT4/INT8/FP8/GGUF) is community, not official. Weights about 49.25 GiB (orcarouter basis). Small teams borrow cloud GPUs or wait for quantization.

Cost shape cannot be compared; each column states its basis. Local runs cost GPU depreciation plus power, low marginal but high upfront; Midjourney and Canva are monthly; Jimeng is membership plus credits; Nano Banana 2 bills per image (from $0.067 per 1024-squared, vendor basis, not retested), zero upfront but costly at scale. Comparing them directly is meaningless, so this piece states basis only.

# Ming-Image inference commands (from the official README, directly citable)
export CUDA_VISIBLE_DEVICES=0
git clone https://github.com/inclusionAI/Ming-Image
cd Ming-Image && pip install -r requirements.txt
# Text-to-design: steps 12, CFG 1.0, recommended resolution 2048
python infer.py --model inclusionAI/Ming-Image-0.1-Design --task text-to-image \
  --prompt "your prompt" --resolution 2048 --output-dir outputs/t2i
# Layer decomposition: steps 12, CFG 2.0, recommended resolution 1024, emits 2-9 RGBA layers
python infer.py --model inclusionAI/Ming-Image-0.1-Design-Layer --task layer \
  --prompt "Decompose this image into N layers ..." --resolution 1024

Vendor speed notes: ai-bot claims roughly 20 seconds per image after engineering optimization and edit cost only 1/7 of GPT-image2; the ai-bot comparison table claims Ming-Image-Layer 6B against Qwen-Image-Layered 20B, with 12 of 12 Crello-Test metrics won and single inference 183s versus 795s (about 4.3x faster). All are tagged "vendor or review basis, not independently retested"; the leaderboard and Crello numbers appear in the model card as charts only, with no citable figures, and this site did not independently retest them.

Who Should Use What

Compress the previous five sections into one division table, and pick directly by role.

Closing thought: a design tool competes not on whose output is more stunning, but on whether you can still control the image after it is made. Of the six, only Ming-Image makes editable layers native and pairs it with MIT and self-hosting, separating it from closed flat output; Qwen-Image 2.1 has RGBA but is non-commercial bound; the other four suit one-shot generation over production iteration. For price and quality, return to the image model cost comparison; for video, see the AI video workbench comparison. The practical takeaway is to choose by where the asset goes next: if it ships and gets revised, layers win outright; if it is a one-off mood board, any flat generator is good enough.

FAQ

Q1: Is Ming-Image really MIT and can I use it commercially? A1: Yes. The MIT license is confirmed consistent across four sources (GitHub API, datanorth.ai, bittide.aicompass.dev, orcarouter.ai) with commercial use explicitly allowed, self-hosting permitted, and derivative work permitted. This is one of the few facts stated flatly. But note its hardware threshold is a single GPU of at least 80 GiB VRAM; 24 GiB consumer cards have no official guidance.

Q2: Can a design tool without layers enter a production workflow? A2: It can do "one-shot generation" but struggles with "repeated iteration". Midjourney, Jimeng, and Nano Banana all return flat single images, so changing one word often means regenerating the whole thing, which is costly and poorly controllable. If editing image and text is a high-frequency action in your workflow, native RGBA plus automatic layering (such as Ming-Image) is nearly mandatory; if you only need an atmospheric image, flat output is enough.

Q3: What hardware does local deployment of Ming-Image actually need? A3: The model card's minimum validated configuration is a single GPU with at least 80 GiB VRAM at BF16, with weights about 49.25 GiB. 24 GiB consumer cards have no official deployment guidance; community quantization (INT4/INT8/FP8/GGUF) is a community route, not an official endorsement. Small teams either borrow large cloud GPUs or wait for community quantization to mature.

Q4: Which is cheaper, Nano Banana 2 or Ming-Image? A4: They cannot be compared directly. Nano Banana 2 bills per image, from about $0.067 per 1024-squared upward (vendor basis, not independently retested), zero upfront but expensive at scale; Ming-Image is local compute whose upfront is the depreciation plus power of an 80 GiB-plus GPU, low marginal cost but high threshold. Their cost denominators differ, so this piece crowns no "cheapest" and only states the basis.

Q5: Why do Jimeng, Midjourney, and Canva, the three closed-source tools, get no exact prices here? A5: To avoid fabricating numbers. Their subscription and credit prices are official and frequently change; exact tiers are per official announcement, so this piece only tags them with the "per official" basis and gives no unretested hard quote. For price comparison, return to the batch 22 cost comparison.

Sources

• GitHub repository inclusionAI/Ming-Image (README and API snapshot, 2026-09-24: 91 stars / 6 forks / Python / MIT)
• Model card and README: native RGBA VAE, Design-Layer 2 to 9 layers, 8K structured prompt, PPT Skill delivery chain
• ai-bot.cn: Ming-Image speed basis (about 20 seconds per image, edit cost 1/7), Crello comparison (12 of 12 wins, 183s versus 795s, 4.3x faster), all tagged vendor or review basis, not independently retested
• orcarouter.ai: Nano Banana 2 pricing ($0.067 per 1024-squared upward), Qwen-Image 2.1 (7.1B DiT, Qwen Research License non-commercial, native RGBA), all tagged vendor or review basis
• datanorth.ai, bittide.aicompass.dev: four-source MIT confirmation
• Related reading: image model cost comparison, AI video workbench comparison

This article is based on public model cards, READMEs, and vendor or review material (as of September 2026); pricing and speed figures were not independently retested and should be confirmed against each official announcement.