TRANSPARENT BY DESIGN
How we estimate VRAM
These are ESTIMATES, not benchmark measurements. We make the assumptions visible so you can judge whether they match your workload.
Units, headroom & fit status
Memory estimates and capacity comparisons use GiB (1,073,741,824 bytes). File sizes in bytes are divided by this value. Hardware names retain their published labels. A nominal capacity is used as a GiB planning value when exact device bytes are unavailable. Total = subtotal × (1 + safety margin). The margin is 12%.
Green: total including margin fits. Yellow: the subtotal fits without margin, or a lighter configuration fits with margin. Red: neither the selected workload nor the fallback configurations checked by the estimator fit. Yellow suggestions come from the same fit functions used by the calculator; a fallback is not a guarantee of backend support.
Available memory per device is its capacity times its usable fraction. Apple unified memory generally uses a 75% planning budget unless a model-specific capacity is listed. A sharded LLM must satisfy the allocation on each device. An ordinary ComfyUI workflow must fit each active stage on one device. Additional devices can run independent copies when the stages fit.
LLM inference
weights bytes = parameters × bits per weight / 8. Measured GGUF rates come from the llama.cpp Llama 3.1 8B reference file, including its mixed tensor types and metadata. The shared FP16/BF16 option uses the FP16 reference rate. Other checkpoints can differ. NVFP4 counts an 8-bit scale per 16 packed weights. Nominal formats model the packed payload only. Generic Q1–Q7 widths describe storage scenarios and do not specify an available checkpoint or inference kernel. NVFP4 block format ↗
KV bytes = 2 × KV layers × KV heads × head dimension × context × batch × bytes per element. The factor 2 represents keys and values. Grouped-query attention uses KV heads rather than all attention heads. Hybrid models use the recorded KV-cache layer count when available; linear-attention state is approximated within runtime overhead.
runtime GiB = first-GPU overhead + extra-GPU overhead × max(0, GPU count − 1). Inference subtotal is weights + KV cache + runtime. Context is capped at the model limit and the calculator maximum.
Effective weight precision
| Quantization | Bits / weight |
|---|---|
| FP16 / BF16 | 16.0005 bpw · FP16 GGUF reference |
| FP8 | 8 bpw · Nominal bit width |
| INT8 | 8 bpw · Nominal bit width |
| NVFP6 | 6 bpw · Nominal bit width |
| NVFP4 | 4.5 bpw · Includes block scale |
| INT4 | 4 bpw · Nominal bit width |
| Q8_0 | 8.5008 bpw · Measured GGUF |
| Q6_K | 6.5633 bpw · Measured GGUF |
| Q5_K_M | 5.7036 bpw · Measured GGUF |
| Q4_K_M | 4.8944 bpw · Measured GGUF |
| Q3_K_M | 3.996 bpw · Measured GGUF |
| Q2_K | 3.1593 bpw · Measured GGUF |
| Q7 | 7 bpw · Nominal bit width |
| Q6 | 6 bpw · Nominal bit width |
| Q5 | 5 bpw · Nominal bit width |
| Q5_K_S | 5.5704 bpw · Measured GGUF |
| Q4_K_S | 4.6672 bpw · Measured GGUF |
| Q4 | 4 bpw · Nominal bit width |
| Q3_K_L | 4.2979 bpw · Measured GGUF |
| Q3_K_S | 3.6429 bpw · Measured GGUF |
| Q3 | 3 bpw · Nominal bit width |
| Q2 | 2 bpw · Nominal bit width |
| Q1 | 1 bpw · Nominal bit width |
KV cache storage
| Type | Bytes / element |
|---|---|
| FP16 | 2 |
| FP8 | 1 |
| FP4 | 0.5 |
| INT4 | 0.5 |
| NVFP4 | 0.5625 |
| Q8_0 | 1.0625 |
| Q4_0 | 0.5625 |
Quantized KV blocks include a scale value; they are larger than a simple 1-byte or half-byte approximation.
Full fine-tuning, LoRA and QLoRA
Full fine-tuning budgets 18 bytes per parameter for mixed-precision AdamW, including 2 for BF16 weights, 4 for master weights, 4 for gradients and 8 for moments. Activations, runtime and headroom are added separately. Multi-GPU training assumes weights, gradients and optimizer states are sharded. Hugging Face ↗
LoRA freezes BF16 base weights and trains adapters. QLoRA freezes NF4 base weights. Both use the adapter and activation estimates below.
QLoRA base weights use double-quantized NF4 at 4.126953125 bits per weight. Inference quantization and KV controls do not apply in QLoRA mode. We assume gradient checkpointing, BF16 activations and Adam optimizer states for LoRA adapters.
LoRA applies to q, k, v, o, gate, up and down projections. Let h = hidden size, q = attention heads × head dimension, k = KV heads × head dimension, and m = h × 3.5. trainable parameters = rank × layers × [(h + q) + 2(h + k) + (q + h) + 3(h + m)].
adapter + optimizer bytes = trainable parameters × 16, covering BF16 weights, BF16 gradients, an FP32 master copy and two FP32 Adam moments. activation bytes = batch × sequence length × hidden size × 2 × (layers + 16). The activation assumption keeps one BF16 hidden state per layer plus approximately 16 hidden widths for the active layer being recomputed.
Subtotal = NF4 base weights + adapter and optimizer + activations + QLoRA runtime + additional GPU overhead. Actual training memory depends strongly on the implementation, target modules and checkpointing policy.
Shared constants
| Safety margin | 12% |
|---|---|
| LLM runtime overhead | 0.75 GiB |
| Each additional GPU overhead | 0.3 GiB |
| QLoRA runtime overhead | 1.5 GiB |
| QLoRA NF4 bits / weight | 4.126953125 |
| LoRA bytes / trainable parameter | 16 |
| Assumed MLP / hidden ratio | 3.5 |
| Apple usable fraction | 0.75 |
Diffusion, images & video: staged peak
ComfyUI runs a workflow in stages and its memory manager moves idle models to system RAM between them (observed in the SDXL measurement below: Unloaded partially: 3056.88 MB freed
). The GPU therefore needs the largest stage, not the sum of every model:
- Text encode = text encoders on GPU + inference reserve
- Sampling = diffusion model + sampling activations + inference reserve
- VAE decode = VAE + decode activations + inference reserve
total = max(stages) × (1 + 0.12). The result card shows every stage and marks the peak. The inference reserve is ComfyUI’s minimum_inference_memory(), 0.8 GiB (source). Weight memory comes from exact source file bytes divided by 1,073,741,824; CPU text encoder offload contributes zero GPU memory, and requested FP8 text encoders fall back individually to FP16 where no FP8 file is recorded.
Keep all models resident in VRAM (off by default, resident=1 in the link) models ComfyUI’s --highvram-style behaviour: every stage also carries every other model’s weights, so the requirement becomes all weights + the largest activation + reserve.
Activation formulas (ComfyUI’s own memory planner)
Coefficients are not fitted by us: they are the constants ComfyUI uses to decide what to load, read from its source at commit 1e719279f5. Activations are 2-byte (BF16/FP16) even when weights are FP8. Divide each spatial dimension by the VAE compression ratio of the checkpoint (8, 16 or 32). Video uses floor((frames − 1) / temporal compression) + 1 latent frames. Mochi uses 6 and LTX-Video uses 8.
- Sampling:
batch × latent elements × 2 bytes × 0.01 × memory_usage_factorMiB, ComfyUI’s memory_required() at the one-conditioning-at-a-time minimum the sampler falls back to (minimum_memory_required). - VAE decode (per batch item, as ComfyUI sizes decode batches to free memory): SD/SDXL/SD3/Flux
2178 × h × w × 64 × 2bytes (source); Wan 2.1(2200 if ≤4 latent frames else 7000) × h × w × 64 × 2(source); CogVideoX2800 × max(2, (t − 1) × 4 + 1) × h × w × 64 × 2(source) applied to one 2-latent-frame chunk, because ComfyUI’s CogVideoX decoder rolls through the video in chunks (source). - If an untiled decode does not fit, ComfyUI retries with tiled decoding automatically (source), so a VAE-only overflow is yellow with a tiled suggestion ( 64 latent pixels per 2D tile, 256 output pixels per 3D tile).
These are ComfyUI’s planning estimates, which its authors mark as conservative in places. They do not model custom nodes (block swap, WanVideoWrapper), attention kernels without PyTorch SDPA, or fragmentation. Exact source default resolutions are preserved, including those outside the 16-pixel input grid.
Calibration table
| Model | memory_usage_factor | VAE decode formula | Public measurement cross-check |
|---|---|---|---|
| FLUX.1 [dev] | 3.1 | KL VAE |
|
| FLUX.1 [schnell] | 3.1 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Stable Diffusion XL Base 1.0 | 0.8 | KL VAE |
|
| Stable Diffusion 3.5 Large | 1.6 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Wan 2.1 T2V 14B | 2.3042 | Wan 2.1 |
|
| Wan 2.1 T2V 1.3B | 0.6913 | Wan 2.1 |
|
| CogVideoX 5B | 2 | CogVideoX | No public measurement found; ComfyUI planner constants only. |
| SDXL Turbo | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Playground v2 512px Base | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Playground v2 1024px Aesthetic | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Playground v2.5 1024px Aesthetic | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| RealVisXL V4.0 | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| RealVisXL V5.0 | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| DreamShaper XL 1.0 | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| DreamShaper XL v2 Turbo | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Juggernaut XL v9 | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| SDXL-Lightning-1step | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| SDXL-Lightning-2step | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| SDXL-Lightning-4step | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| SDXL-Lightning-8step | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Latent Consistency SDXL | 0.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| FLUX.1 Krea [dev] | 3.1 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Stable Diffusion 3.5 Large Turbo | 1.6 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| Qwen-Image | 1.8 | Wan 2.1 | No public measurement found; ComfyUI planner constants only. |
| Qwen-Image-2512 | 1.8 | Wan 2.1 | No public measurement found; ComfyUI planner constants only. |
| HunyuanImage 2.1 | 8.7 | HunyuanImage 2.1 | No public measurement found; ComfyUI planner constants only. |
| HunyuanImage 2.1 Distilled | 8.7 | HunyuanImage 2.1 | No public measurement found; ComfyUI planner constants only. |
| Z-Image-Turbo | 2.8 | KL VAE | No public measurement found; ComfyUI planner constants only. |
| CogVideoX 2B | 2 | CogVideoX | No public measurement found; ComfyUI planner constants only. |
| Mochi 1 Preview | 2 | Mochi 1 | No public measurement found; ComfyUI planner constants only. |
| HunyuanVideo T2V 720p | 1.8 | HunyuanVideo | No public measurement found; ComfyUI planner constants only. |
| HunyuanVideo 1.5 T2V 480p | 4 | HunyuanVideo 1.5 | No public measurement found; ComfyUI planner constants only. |
| HunyuanVideo 1.5 T2V 720p | 4 | HunyuanVideo 1.5 | No public measurement found; ComfyUI planner constants only. |
| HunyuanVideo 1.5 T2V 480p CFG Distilled | 4 | HunyuanVideo 1.5 | No public measurement found; ComfyUI planner constants only. |
| Wan 2.2 TI2V 5B | 1.3825 | Wan 2.2 | No public measurement found; ComfyUI planner constants only. |
| LTX-Video 0.9.1 (2B) | 5.5 | LTX-Video | No public measurement found; ComfyUI planner constants only. |
| LTX-Video 0.9.5 (2B) | 5.5 | LTX-Video | No public measurement found; ComfyUI planner constants only. |
| LTX-Video 0.9.7-dev (13B) | 11 | LTX-Video | No public measurement found; ComfyUI planner constants only. |
| LTX-Video 0.9.7-distilled (13B) | 11 | LTX-Video | No public measurement found; ComfyUI planner constants only. |
| LTX-Video 0.9.8-13B-distilled (13B) | 11 | LTX-Video | No public measurement found; ComfyUI planner constants only. |
Data sources & limitations
Architecture values come from Hugging Face config.json files and parameter counts from the Hugging Face API safetensors metadata. Component weight sizes come from file listings. Each model guide links its exact sources and last verification date.
- Hugging Face configuration reference
- Hugging Face Hub API
- llama.cpp quantize README: effective bits per weight
VRAM estimates run locally. The directory uses dated specifications and prices and can refresh the trending list from OpenRouter. A memory verdict is not a throughput, compatibility or hardware purchase recommendation. Verify actual usage with your target backend and leave room for other applications, the display and system allocations.