JD Open Source dropped JoyAI-Image-Edit-Plus on 2026-06-23, and by mid-July it was already merged into the Diffusers pipeline. The model is a 24B unified system — an 8B Qwen-VL-class MLLM stapled to a 16B MMDiT — trained specifically for instruction-guided local edits (“remove the person on the left”, “make the sky sunset”). That extra MLLM tower is the whole story for buyers. It is not swappable, it does not offload cleanly, and it changes what “enough VRAM” means.
Quick answer: The RTX 4090 (24GB) is the best GPU for JoyAI Image Edit if you want BF16 quality without babysitting memory. If FP8 is fine, the RTX 4070 Ti Super (16GB) runs the same workflow at roughly ~13GB active with a small quality trade-off.
NVIDIA GeForce RTX 4090
24GB GDDR6X24GB VRAM covers the full 24B model at BF16 without offloading — MLLM tower and MMDiT tower both resident. Enough headroom for batched edits and IP-Adapter-style guidance on top.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Who this guide is for
This is a scenario-based guide for image editing, not image generation. If you are generating from scratch — SDXL, Flux, Qwen Image — the calculus is different, and our Qwen Image ranking is the closer match. JoyAI-Image-Edit-Plus is a different beast: you feed it a source image plus a natural-language instruction, and it returns a locally-edited result. Think Photoshop generative fill, but locally hosted and instruction-driven.
The people who should read this guide:
- Photographers batching background swaps or object removal locally
- ComfyUI workflow builders wiring JoyAI as an edit node between generation and upscale
- Anyone running Flux Kontext today who wants to A/B against JoyAI’s edit quality
- Teams that cannot send client images to hosted APIs
If you are already comfortable running a broader AI photo editing rig, the shift to JoyAI is a VRAM story — everything else you already know.
Why edit-workflow VRAM is different
Here is the trap most buyers fall into. They see “24B parameters” and mentally compare to a 24B LLM, which fits comfortably at 4-bit on a 16GB card. JoyAI does not work that way.
The 24B is two towers that must be resident at the same time:
- 8B Qwen-VL-class MLLM — parses the instruction against the source image, holds visual tokens in memory across the full edit
- 16B MMDiT — the diffusion transformer that actually renders the edit
You cannot page one out mid-edit. The MLLM re-references the source through cross-attention while the MMDiT denoises, so both towers stay hot for the entire step loop. That means VRAM math for JoyAI is closer to “sum of both towers plus activations” than to a swappable LLM setup.
The practical consequence: JoyAI feels heavier than its parameter count suggests, and cards that comfortably run 12B Flux.1 Dev — or 7B Qwen-Image-Edit — will not run JoyAI at native precision. This is the single most important thing to internalize before spending money.
JoyAI Image Edit VRAM requirements
Rough resident-memory budget for a single 1024px edit at 20 steps:
| Precision | Active VRAM | Minimum card | Quality | Notes |
|---|---|---|---|---|
| FP16 (native) | ~28GB | 32GB (5090) | Reference | Cleanest results, only fits on 5090 or A6000-class |
| BF16 | ~24-26GB | 24GB (4090) | ~99% of FP16 | The realistic quality target for local users |
| FP8 (e4m3fn) | ~13GB | 16GB (4070 Ti Super) | ~95-97% | Slight softness on fine text and hair details |
| INT8 (weight-only) | ~11GB | 12GB (3060/4070) | ~90-93% | Noticeable artifacts on complex edits |
| Q4 attempt | ~8GB | 10GB+ | Broken | Do not bother — the MLLM tower degrades badly below FP8 |
The BF16 line is the one that matters. The Diffusers pipeline defaults to BF16 on Ampere and newer, and 24GB is the minimum for it to fit without CPU offload. FP8 shifts the floor to 16GB, and that is where the price/quality curve gets interesting.
For a broader primer on what these numbers mean in practice, how much VRAM for Stable Diffusion covers the underlying diffusion memory model.
JoyAI vs Qwen-Image-Edit vs Flux Kontext
Rough VRAM footprint for the three obvious competitors in the instruction-edit space:
| Model | Params | BF16 VRAM | FP8 VRAM | Best for |
|---|---|---|---|---|
| Qwen-Image-Edit | ~7B | ~14GB | ~7GB | 12-16GB cards, fastest iteration |
| Flux Kontext | ~12B | ~22GB | ~12GB | Kontext-style multi-image reasoning |
| JoyAI-Image-Edit-Plus | 24B (8B + 16B) | ~24-26GB | ~13GB | Highest edit fidelity, 24GB cards |
Here is my contrarian take, because I think buyers need it: Qwen-Image-Edit at 7B may be a better fit for 12-16GB cards. JoyAI is only worth it if you already have 24GB or you are willing to accept FP8. On paper JoyAI wins the fidelity crown, in real workflows the 3-4x speed advantage of Qwen-Image-Edit at FP16 on the same 16GB card often matters more. Do not chase 24B just because it is newer.
If you are Flux-native already, our Flux.2 GPU guide covers the family in more depth — Kontext sits closer to Flux than to JoyAI in workflow feel.
5 GPUs ranked for JoyAI Image Edit
1. RTX 5090 (32GB, ~$4,900) — FP16 native, if you must
The 5090 is the only consumer card that runs JoyAI at native FP16 without offloading. 32GB covers both towers plus batch-of-2 headroom. Blackwell’s FP8 kernels also give a real 25-30% speedup versus the 4090 on the MMDiT tower specifically.
The catch is you are paying $400+ over the 4090 mostly for FP16 quality that is a hair above BF16. Unless you are producing client work where the last 1% of fidelity is billable, the 5090 is overkill for JoyAI alone.
2. RTX 4090 (24GB, ~$2,200) — the correct default
This is the card I recommend most JoyAI Image Edit buyers actually buy. 24GB is exactly enough for BF16 with both towers resident, no offloading, no gymnastics. Expect ~6-8 seconds per 1024px edit at 20 steps in Diffusers, roughly ~4-5 seconds with FP8 for iteration.
Check NVIDIA GeForce RTX 4090 on Amazon→Buy on Shopee SG→The 4090 also stays useful for everything else in an editing rig — ComfyUI workflows, Flux, Qwen Image, LoRA training. If you are building a serious local edit pipeline, the 4090 is the anchor.
3. RTX 5080 (16GB, ~$1,400) — FP8 with modern architecture
The 5080 is the Blackwell FP8 pick for people who do not want the 4090’s price. 16GB is enough for JoyAI at FP8 (~13GB active), with a bit of headroom for a single control image or IP-Adapter reference. Expect ~5-7 seconds per edit at FP8.
Where the 5080 breaks down is BF16 — it does not fit, full stop. If you are willing to live at FP8 forever, the 5080 is a clean pick. If you might want BF16 quality later, do not settle for 16GB.
4. RTX 5070 Ti (16GB, ~$1,050) — Blackwell value
Same 16GB envelope as the 5080, ~20% slower on the MMDiT tower, but $250 cheaper. For a JoyAI-focused build at FP8, the 5070 Ti is the sensible Blackwell pick. Roughly ~7-9 seconds per 1024px edit at FP8, which is fine for interactive iteration but not for high-throughput batch work.
5. RTX 4070 Ti Super (16GB, ~$800) — the FP8 sweet spot
The 4070 Ti Super is the value FP8 pick. Ada Lovelace’s FP8 kernels are mature, 16GB fits JoyAI at FP8 comfortably, and ~$700 street price undercuts every Blackwell option. Expect ~8-10 seconds per edit at FP8, no BF16 possible.
Check NVIDIA GeForce RTX 4070 Ti Super on Amazon→Buy on Shopee SG→Honest note: if I were buying today with a $700 budget and I only needed edit workflows, I would strongly consider skipping JoyAI entirely and running Qwen-Image-Edit at FP16 on this same card. Faster, cleaner, same 16GB envelope.
Runner-up: RTX 4060 Ti 16GB (~$425)
Technically it fits JoyAI at FP8. Practically, at ~22-28 seconds per edit, iteration feels miserable. If your budget is this tight, use cloud until you can step up.
Not sure yet? Rent one first
Buying blind for a 24B edit model is a mistake. RunPod runs a 4090 for about $0.50/hr — you can validate JoyAI fits your actual workflow (batch size, control images, throughput) in an evening for less than $10.
Rent an RTX 4090 on RunPod to test JoyAI Image Edit→Which GPU should YOU buy?
Buy the RTX 4090 if:
- You want BF16 quality (the realistic max on consumer cards)
- You edit for clients or produce work where fidelity matters
- You also run Flux Kontext, Flux.2, or LoRA training
Buy the RTX 5090 if:
- You need FP16 reference quality and batched edits
- JoyAI is one of several 20B+ models you run daily
- Budget is not the constraint
Buy the RTX 4070 Ti Super or RTX 5070 Ti if:
- FP8 quality is acceptable to you (it is, for most work)
- You want the cheapest sane entry into JoyAI
- You accept you will never run BF16 on this card
Buy Qwen-Image-Edit instead of JoyAI if:
- Your card has 12-16GB and speed matters more than the last 3% of edit fidelity
- You want FP16 not FP8
Skip:
- Any 8-10GB card — JoyAI simply does not run well at Q4 or below
Common mistakes to avoid
- Conflating generation VRAM and edit VRAM. A card that runs Flux.1 Dev generation at 24GB does not automatically run JoyAI edits comfortably. The MLLM tower adds ~6-8GB of resident memory that generation-only pipelines do not carry.
- Loading BF16 on a 16GB card. The Diffusers pipeline will happily try, then OOM mid-step, or silently CPU-offload the MLLM tower and drop your speed by 5-10x. If you are on 16GB, force FP8 explicitly.
- Not batching edits. JoyAI’s setup cost (loading both towers, warming the MLLM) is high. Running 20 edits one-at-a-time is 3-5x slower than batching them. Always queue.
- Ignoring the MLLM tower in memory math. Every VRAM estimate I have seen online quotes only the 16B MMDiT. The 8B MLLM is real, it stays resident, and it turns “fits 16GB” into “does not fit 16GB” if you forget it.
Final verdict
| Budget | GPU | JoyAI capability |
|---|---|---|
| ~$425 | RTX 4060 Ti 16GB | FP8, slow (~25 s/edit) |
| ~$800 | RTX 4070 Ti Super | FP8, ~8-10 s/edit |
| ~$1,050 | RTX 5070 Ti | Blackwell FP8, ~7-9 s/edit |
| ~$1,400 | RTX 5080 | Fastest 16GB FP8 card |
| ~$2,200 | RTX 4090 | BF16 native, ~6-8 s/edit |
| ~$4,900+ | RTX 5090 | FP16 reference quality |
NVIDIA GeForce RTX 4090
24GB GDDR6X24GB fits both towers at BF16 without offloading. If you edit locally and care about fidelity, this is the card. Also stays useful for Flux Kontext, Qwen-Image-Edit, and LoRA training in the same rig.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
For most JoyAI Image Edit buyers, the RTX 4090 is the correct answer. Step down to the 4070 Ti Super only if FP8 quality is genuinely acceptable and budget forces the trade — and if you land there, seriously reconsider whether Qwen-Image-Edit at FP16 is the better local model for your card.
JoyAI is a 24B unified system, not a 16B diffusion model. Buy VRAM accordingly — or buy Qwen-Image-Edit.
Frequently asked questions
How much VRAM does JoyAI Image Edit need?
JoyAI-Image-Edit-Plus needs roughly 24-26GB at BF16 and about 13GB at FP8 for a single 1024px edit. The model is a 24B unified system with an 8B MLLM tower and a 16B MMDiT tower that both stay resident during the edit, so VRAM math is closer to the sum of both towers than a swappable LLM setup.
Can a 16GB GPU run JoyAI Image Edit?
Yes, at FP8 quantization. Cards like the RTX 4070 Ti Super, RTX 5070 Ti, and RTX 5080 run JoyAI at FP8 with roughly 13GB active, taking around 8-10 seconds per 1024px edit on Ada Lovelace. BF16 does not fit in 16GB without CPU offloading, which slows the pipeline dramatically.
Is JoyAI Image Edit better than Qwen-Image-Edit?
JoyAI produces higher-fidelity edits than Qwen-Image-Edit on complex instructions, but it is roughly 3x heavier at 24B parameters versus about 7B. On a 16GB card, Qwen-Image-Edit runs at native FP16 and iterates 2-3x faster, which often outweighs the fidelity gap for interactive workflows. Choose JoyAI if you have 24GB and edit for clients; choose Qwen-Image-Edit for 12-16GB cards.
Does JoyAI Image Edit work in Diffusers?
Yes. The JoyAI-Image-Edit-Plus pipeline was merged into the Hugging Face Diffusers library in May 2026 and works out of the box on Ampere, Ada Lovelace, and Blackwell GPUs. The pipeline defaults to BF16 on 24GB cards and supports an FP8 mode for 16GB cards. Model weights are hosted on Hugging Face as jdopensource/JoyAI-Image-Edit.
Should I buy the RTX 4090 or wait for a cheaper 24GB option for JoyAI?
The RTX 4090 remains the correct 24GB pick for JoyAI Image Edit through 2026. It is the only consumer card at 24GB that runs BF16 without offloading, and no 24GB Blackwell equivalent exists below the 5090. Used RTX 3090 24GB is a value alternative at around $820, roughly 2x slower on the MMDiT tower but functionally identical on capability.