iris-3b-int8 / benchmark.json
Aikimi's picture
Add validated Iris-3B weight-only INT8 models and CUDA comparisons
6231647 verified
Raw History Blame Contribute Delete
9.22 kB
{
"date": "2026-10-09",
"environment": {
"torch": "2.13.0+cu130",
"python": "3.13.14 (main, Jun 11 2026, 04:04:46)",
"device": "NVIDIA GeForce RTX 3090",
"cuda": "13.0",
"torch_threads": 10
},
"method": "One GPU; one task model at a time; torch inference mode and BF16 autocast. T2I uses Qwen3-VL-4B and 100 steps, 1024x1024, CFG 3 after a 2-step warmup. Other tasks have no warmup; their first photo case includes cold GPU transfer and host memory/disk effects. Times are Runner.run wall time including GPU offload, excluding initial CPU model construction and downloads. Initial 2-step T2I cold-load smoke is excluded. Each condition was measured once, not a statistical speed benchmark. Peak is torch CUDA allocated/reserved memory for the inference. It excludes other processes and is not total board memory. Depth uses the 1024x1024 normal T2I PNG. Restoration inputs are generated references downsampled to 256x256 JPEG quality 35 (one tile) and cropped/downsampled to 384x256 JPEG quality 45 (two overlapping tiles), outputs 1024x1024 and 1536x1024.",
"records": [
{
"task": "generate",
"case": "photo",
"precision": "normal",
"request": {
"task": "generate",
"precision": "normal",
"prompt": "A candid editorial photograph of a young adult woman with short dark hair standing at a rain soaked Tokyo street crossing at dusk, holding a transparent umbrella, reflections of neon signs, realistic skin texture, 50mm lens, natural proportions, cinematic but believable lighting.",
"seed": 424242,
"width": 1024,
"height": 1024,
"steps": 100,
"cfg": 3.0,
"negative_prompt": ""
},
"size": [
1024,
1024
],
"wall_seconds": 243.98675999999978,
"peak_allocated_mib": 19348.044921875,
"peak_reserved_mib": 19732.0,
"pixel_std": 72.87049865722656
},
{
"task": "generate",
"case": "photo",
"precision": "int8",
"request": {
"task": "generate",
"precision": "int8",
"prompt": "A candid editorial photograph of a young adult woman with short dark hair standing at a rain soaked Tokyo street crossing at dusk, holding a transparent umbrella, reflections of neon signs, realistic skin texture, 50mm lens, natural proportions, cinematic but believable lighting.",
"seed": 424242,
"width": 1024,
"height": 1024,
"steps": 100,
"cfg": 3.0,
"negative_prompt": ""
},
"size": [
1024,
1024
],
"wall_seconds": 244.0833332000766,
"peak_allocated_mib": 10934.3515625,
"peak_reserved_mib": 11488.0,
"pixel_std": 73.21208953857422
},
{
"task": "generate",
"case": "watercolor",
"precision": "normal",
"request": {
"task": "generate",
"precision": "normal",
"prompt": "A detailed hand painted watercolor illustration of a small wooden reading room above a quiet bookshop. A ginger cat sleeps on a green velvet armchair, warm light from a brass desk lamp, a large window looking onto rainy rooftops, bookshelves with distinct books, delicate line work, paper texture. No text.",
"seed": 20261009,
"width": 1024,
"height": 1024,
"steps": 100,
"cfg": 3.0,
"negative_prompt": ""
},
"size": [
1024,
1024
],
"wall_seconds": 252.1166963998694,
"peak_allocated_mib": 19347.509765625,
"peak_reserved_mib": 19752.0,
"pixel_std": 66.97412872314453
},
{
"task": "generate",
"case": "watercolor",
"precision": "int8",
"request": {
"task": "generate",
"precision": "int8",
"prompt": "A detailed hand painted watercolor illustration of a small wooden reading room above a quiet bookshop. A ginger cat sleeps on a green velvet armchair, warm light from a brass desk lamp, a large window looking onto rainy rooftops, bookshelves with distinct books, delicate line work, paper texture. No text.",
"seed": 20261009,
"width": 1024,
"height": 1024,
"steps": 100,
"cfg": 3.0,
"negative_prompt": ""
},
"size": [
1024,
1024
],
"wall_seconds": 254.52246790006757,
"peak_allocated_mib": 10933.671875,
"peak_reserved_mib": 11528.0,
"pixel_std": 67.01991271972656
},
{
"task": "depth",
"case": "photo",
"precision": "normal",
"request": {
"task": "depth",
"precision": "normal"
},
"size": [
1024,
1024
],
"wall_seconds": 283.81315309996717,
"peak_allocated_mib": 15374.9541015625,
"peak_reserved_mib": 16048.0,
"pixel_std": 85.49915313720703
},
{
"task": "depth",
"case": "photo",
"precision": "int8",
"request": {
"task": "depth",
"precision": "int8"
},
"size": [
1024,
1024
],
"wall_seconds": 87.89453630009666,
"peak_allocated_mib": 6960.0576171875,
"peak_reserved_mib": 7674.0,
"pixel_std": 85.4561996459961
},
{
"task": "depth",
"case": "watercolor",
"precision": "normal",
"request": {
"task": "depth",
"precision": "normal"
},
"size": [
1024,
1024
],
"wall_seconds": 8.675420299870893,
"peak_allocated_mib": 15406.9013671875,
"peak_reserved_mib": 16058.0,
"pixel_std": 80.67788696289062
},
{
"task": "depth",
"case": "watercolor",
"precision": "int8",
"request": {
"task": "depth",
"precision": "int8"
},
"size": [
1024,
1024
],
"wall_seconds": 3.3528326998930424,
"peak_allocated_mib": 6989.7626953125,
"peak_reserved_mib": 7708.0,
"pixel_std": 80.69136047363281
},
{
"task": "upscale",
"case": "photo",
"precision": "normal",
"request": {
"task": "upscale",
"precision": "normal"
},
"size": [
1024,
1024
],
"wall_seconds": 300.6608500001021,
"peak_allocated_mib": 15366.005859375,
"peak_reserved_mib": 16018.0,
"pixel_std": 74.8482437133789
},
{
"task": "upscale",
"case": "photo",
"precision": "int8",
"request": {
"task": "upscale",
"precision": "int8"
},
"size": [
1024,
1024
],
"wall_seconds": 91.08279359992594,
"peak_allocated_mib": 6953.5810546875,
"peak_reserved_mib": 7654.0,
"pixel_std": 74.83624267578125
},
{
"task": "upscale",
"case": "watercolor",
"precision": "normal",
"request": {
"task": "upscale",
"precision": "normal"
},
"size": [
1536,
1024
],
"wall_seconds": 9.50692670000717,
"peak_allocated_mib": 15452.650390625,
"peak_reserved_mib": 16100.0,
"pixel_std": 72.71382141113281
},
{
"task": "upscale",
"case": "watercolor",
"precision": "int8",
"request": {
"task": "upscale",
"precision": "int8"
},
"size": [
1536,
1024
],
"wall_seconds": 4.555513399885967,
"peak_allocated_mib": 7039.1875,
"peak_reserved_mib": 7754.0,
"pixel_std": 72.731201171875
}
],
"comparisons": [
{
"task": "generate",
"case": "photo",
"pixel_mae_0_255": 2.3095102310180664,
"identical_pixels_fraction": 0.023015975952148438
},
{
"task": "generate",
"case": "watercolor",
"pixel_mae_0_255": 4.086854457855225,
"identical_pixels_fraction": 0.0079193115234375
},
{
"task": "depth",
"case": "photo",
"pixel_mae_0_255": 0.5848824381828308,
"identical_pixels_fraction": 0.5262269973754883,
"depth_pearson": 0.9999082843818213,
"depth_shapes": [
[
1024,
1024
],
[
1024,
1024
]
]
},
{
"task": "depth",
"case": "watercolor",
"pixel_mae_0_255": 0.6243413090705872,
"identical_pixels_fraction": 0.5460033416748047,
"depth_pearson": 0.9999447206409359,
"depth_shapes": [
[
1024,
1024
],
[
1024,
1024
]
]
},
{
"task": "upscale",
"case": "photo",
"pixel_mae_0_255": 0.6896772384643555,
"identical_pixels_fraction": 0.09531784057617188
},
{
"task": "upscale",
"case": "watercolor",
"pixel_mae_0_255": 1.0330185890197754,
"identical_pixels_fraction": 0.07715797424316406
}
]
}