SEEN2D S1 · 在 SEEN VAE 上蒸馏 FLUX.2 klein-4B（T2I @512）

job t-20261104024324-zgwfc · 8N · /deemos-research-new/3DLM/saving/voxture/ldm/SEEN2D_S1_3840x25_KleinDistill_512_8N · 每 2 分钟自动刷新 · 总看板 7870 · TensorBoard 6006
step1,320 / 200,0000.66%
速度4.04 s/步最近 15 次日志均值
剩余9天7时按当前速度到 total
loss_fm1.3031起点 2.187
显存峰值33.8 GiB每 20 步重置
最后日志11-03 20:15 UTC40 秒前

训练曲线

flow-matching loss（每 20 步均值，EMA 0.9 平滑）
1.2551.5891.9222.256203456709951,320
REPA loss（1 − cos 对 DINOv3，EMA 0.9 平滑）
0.22720.5060.78481.064203456709951,320
eval：学生 LPIPS（越低越好）
对 K̃（Klein 经 SEEN 往返，验收口径）对 Klein 原图
0.7660.76990.77390.77780
速度 s/步（含 Klein 生成）
3.7524.3474.9435.538203456709951,320
学习率
0.0e+002.4e-054.8e-057.1e-05203456709951,320
文本行数（batch 内最长 prompt，64 对齐）
236257278299203456709951,320
每个生成样本被用的次数
2.2772.8113.3453.879203456709951,320
EMA decay
0.66540.78280.90011.018203456709951,320

In-loop eval（EMA，每 2000 步，128 条 PartiPrompts/GenEval，固定噪声）

验收口径是「学生 vs K̃」：K̃ 是 Klein 的图经 SEEN VAE 编解码后的样子，即这个 VAE 下能达到的上限。前 4000 步的 EMA 仍带不少初始化成分，不要读。
stepLPIPS 对 K̃LPIPS 对 KleinK̃ 对 Klein PSNRprompt 数rank 数EMA 恢复耗时 s
00.76680.777032.4712864✓11
step 0　每行：Klein 原图 | K̃（SEEN 往返=上限） | 学生 EMA（28 步，CFG 3）
  1. a hot air balloon
  2. a boy and a tiger
  3. A photograph of the inside of a subway train. There are red pandas sitting on the seats. One of them is reading a newspaper. The window shows the jungle in the
  4. a portrait of a statue of the Egyptian god Anubis wearing aviator goggles, white t-shirt and leather jacket. the skyline of Los Angeles at night can be seen in

在线生成抽检（Klein 现场给学生出的训练图，rank0 每 50 轮存一次，最新在前）

step0001200_r000400 · token 网格 [37, 28]
  1. flux_reasonWhimsically suspended from a gnarled tree branch, a colossal blue whale rests peacefully in a wooden cradle, its body cradled gently by sturdy ropes that dangle from the dark, leafy boughs above. The serene aquatic giant
  2. pixmo_capThe image depicts a tiled surface, possibly consisting of wet concrete, laid out in a checkerboard pattern with alternating colors of red, white, brown, yellow, and a few other hues such as an off-green, rust, and a touc
  3. flux_reasonSerenity and introspection blend as a solitary figure stands on the precipice, gazing into the boundless expanse of azure mountains, where the whispers of the wind carry the promise of freedom and the weight of infinite
  4. flux_reasonLuxurious high-performance seats dominate the interior of a 2015 Bentley Continental GT GT3-R, captured in Greenwich, CT, showcasing a striking blend of elegance and sportiness. The seats, upholstered in black leather wi
step0001050_r000350 · token 网格 [39, 26]
  1. pixmo_capThis image portrays a framed digital artwork, likely created and rendered on a computer, of two dolphins swimming in the water. The frame is elaborate, featuring an exterior layer in brown, an interior layer in gold, and
  2. flux_reasonRadiant joy and timeless elegance intertwine as two souls celebrate their union, their smiles echoing the promise of love and new beginnings.
  3. flux_reasonElegantly draped across a white-framed window, a set of cream-colored curtains exudes a touch of classic sophistication. The topmost layer features a valance with a soft, gathered design, adorned with subtle gold accents
  4. mj_v6_recapPerson in a parked motor bike Airplane attendants are standing outside a house
step0000900_r000300 · token 网格 [43, 24]
  1. flux_reasonOlivia Watson, an Austin, Texas-based fashion lifestyle blogger, exudes effortless charm and style in this vibrant outdoor portrait set against a lush, garden-like backdrop. She stands confidently amidst blooming pink fl
  2. flux_reasonSerenity whispers through the still lake, where a dog's patient gaze and the gentle sway of trees invite reflection on the quiet beauty of solitude and the enduring promise of nature’s embrace.
  3. flux_reasonComparison showcasing two iconic smartphones: a PalmOne Treo with its classic interface featuring a stylus-friendly design and a grid of colorful icons, positioned to the left, alongside an iPhone with its sleek black fr
  4. flux_reasonA sleek 1966 Ford Fairlane, showcased in a polished dark blue finish, dominates the foreground of this image, capturing the essence of classic American automotive design. The car's iconic features, including its chrome g
step0000750_r000250 · token 网格 [24, 43]
  1. flux_reasonTwo volunteers stand together in a well-lit pantry, wearing aprons that read "St. Francis Soup Kitchen," surrounded by shelves stocked with canned goods and festive decorations, embodying their commitment to community se
  2. flux_reasonEmerging from jagged, irregular stone walls, a wooden bookshelf stands as a surreal fusion of nature and craftsmanship, its four levels cradling an array of colorful books that seem to defy gravity. The stone facade, rou
  3. pd12mThe image shows a black and yellow garden spider perched on the side of a building, with an object attached to the wall in the background.
  4. pd12mThe image shows a small siskin perched atop a tree branch surrounded by lush green leaves and a bright blue sky in the background.
step0000600_r000200 · token 网格 [26, 39]
  1. flux_reasonMahon Pool in Sydney offers a serene escape with its calm waters reflecting the tranquil sky and distant cityscape.
  2. pixmo_capThis color photograph provides a close-up view of the interior of an outdoor grill featuring three large pieces of meat, possibly ribs, that are beautifully grilled. The grill appears to have two levels: two pieces of me
  3. pd12mThe image shows a group of men standing next to each other in an airport. They are all wearing suits and some of them are holding bags. The image is in black and white and there are glass windows in the background.
  4. mj_v6_recapA bed with a clock tower standing tall in a black food truck
step0000450_r000150 · token 网格 [43, 24]
  1. pixmo_capThe image showcases a 3D character design that transitions from an unrendered CAD model to fully rendered illustrations, incorporating elements of both sci-fi and fantasy. This small alien baby, bathed in shades of blue
  2. flux_reasonBad Lack Machine's striking gold and navy-blue branded canopy tent stands prominently outdoors, set against a serene backdrop of lush greenery and a tranquil body of water. The tent features a custom-printed top with bol
  3. flux_reasonDusty red dirt road stretches straight toward distant mesas under a clear blue sky, flanked by low green shrubs and grasses at Ghost Ranch, New Mexico.
  4. flux_reason一位自信的男人站在一辆经典绿色复古车前，穿着一件鲜艳的红色夹克，内搭一件干净的白色衬衫，搭配米色裤子和棕色皮鞋，展现出一种时尚而优雅的风格。
step0000300_r000100 · token 网格 [26, 39]
  1. pd12mThe image shows a painting of a crowd of people in a wooded area, with trees, houses, animals, and other objects in the foreground, and a sky with clouds in the background.
  2. flux_reasonPositioned prominently against a pristine white backdrop, the Trek EVO EX 8 Full Suspension Mountain Bike in KT20 exudes a sleek and robust design, ideal for off-road adventures. The bike features a striking black frame
  3. flux_reasonDelicate gingerbread cookies, intricately decorated with white icing swirls, festive red and green sprinkles, and a green star-shaped cookie topped with three pecans, create a charming holiday display.
  4. flux_reasonSet against a backdrop of lush green trees and a clear blue sky, this charming narrow lot house exudes a classic and inviting aesthetic. The structure features a symmetrical design with a prominent front garage on the ri
step0000150_r000050 · token 网格 [43, 24]
  1. pixmo_capThe photograph captures a minimalist and well-designed section of a room. Prominently featured is a fluffy white chair with angled wooden legs, draped gracefully with a high-quality army green throw. Adjacent to the chai
  2. flux_reasonColorful donuts arranged neatly in a box display a vibrant assortment of flavors and toppings, featuring three rows of four donuts each, with glazes in shades of yellow, chocolate, pink, and white, adorned with colorful
  3. pd12mThe image shows a blackbird perched atop a tree stump in the woods, surrounded by lush green leaves in the background.
  4. mj_v6_recapUrban street with people and traffic