[{"data":1,"prerenderedAt":181},["ShallowReactive",2],{"project-38255201":3},{"id":4,"name":5,"line":5,"line_source":6,"title":7,"url":8,"domain":9,"maker":10,"posted_at":11,"year":12,"points":13,"comments":13,"rank_in_year":14,"shown_in_year":15,"top":16,"tier":17,"category":18,"subcategory":19,"format":20,"keywords":21,"reception":25,"sentiment":26,"stance_mix":27,"scored_comments":29,"aspects":30,"strip":33,"post_text":26,"what":35,"said":58,"maker_in_thread":107,"shape":119,"grid":120,"attempts":145,"same_week":146,"maker_shows":108,"maker_hidden":174,"reader_version":175,"card":176},"38255201","Realtime LCM tool","title","Show HN: Realtime LCM tool","https:\u002F\u002Ftwitter.com\u002FTitusTeatus\u002Fstatus\u002F1723961348101677353","twitter.com","dvrp",1699908626,2023,12,2206,3109,71,"shown","ai","realtime image generation","web_app",[19,22,23,24],"latent consistency model","stable diffusion","gpu inference","quiet",null,{"positive":28,"neutral":29,"mixed":28,"critical":28},0,4,{"performance":31,"idea":32},{"positive":28,"critical":28},{"positive":28,"critical":28},[28,34,28,28,28,34,28,28,34,28,34,28],3,{"problem":26,"audience":26,"pricing":26,"motivation":26,"time_to_build":26,"stack":36,"openness":55,"stage":56,"team_size":57},[37,45],{"value":38,"source":39,"evidence":40},"PyTorch","commenter",[41],{"id":42,"by":43,"text":44},"38255533","brucethemoose2","Are y'all running the diffusion in PyTorch eager mode? AITemplate, stable-fast, or even a torch.compile could get that down to 60ms, I bet. Though I'm not sure the example…",{"value":46,"source":39,"evidence":47},"Stable Diffusion UNet architecture",[48,52],{"id":49,"by":50,"text":51},"38255622","erwannmillon","LCMs are actually exactly SD architecture. LCM is initialized from a regular SD unet and finetuned on a new objective. We are already compiling to get to these times. A lot of…",{"id":53,"by":50,"text":54},"38255346","i worked on the gpu\u002Finfra side of this, so feel free to AMA Ultimately the LCM is just a SD Unet trained with a new objective, so a lot of SD optimizations are transferable to LCMs","unknown","beta","team",{"praise":59,"objections":64,"requests":75,"questions":83,"alternatives":95},[60],{"value":61,"source":39,"evidence":62},"Fast inference times noted as impressive",[63],{"id":42,"by":43,"text":44},[65,71],{"value":66,"source":39,"evidence":67},"Claimed 50-100ms speeds may rely on quality-degrading methods",[68],{"id":69,"by":43,"text":70},"38256890","They might be \"cheating\" with methods that significantly degrade quality.",{"value":72,"source":39,"evidence":73},"Could use fewer inference steps elsewhere at quality cost",[74],{"id":49,"by":50,"text":51},[76],{"value":77,"source":39,"evidence":78},"Support for custom images",[79],{"id":80,"by":81,"text":82},"38255240","vmatsiiako","Can you use custom images?",[84,88],{"value":85,"source":39,"evidence":86},"Is it running in PyTorch eager mode?",[87],{"id":42,"by":43,"text":44},{"value":89,"source":39,"evidence":90},"Is it running on H100s or A100s?",[91],{"id":92,"by":93,"text":94},"38255530","ilaksh","Is it running on H100s or A100s now?",[96,103],{"value":97,"source":39,"evidence":98},"stable-fast",[99,102],{"id":100,"by":43,"text":101},"38256600","> already compiling Hmm, well if you mean torch.compile, y'all should still check out stable-fast, which is claiming ~16ms\u002Fiter on a 4090:…",{"id":42,"by":43,"text":44},{"value":104,"source":39,"evidence":105},"AITemplate",[106],{"id":42,"by":43,"text":44},{"comments":108,"answered":34,"key":109},5,[110,113,116],{"id":111,"text":112},"38255651","some people even claim they get ~50ms on (powerful) consumer hardware (e.g. RTX 4090). we still have not been able to replicate those results; also because we want to do it in a…",{"id":114,"text":115},"38255361","i also worked a bit on the infrastructure! there's a very cool consistent hashing and probability stuff going on for the GPU routing logic but again, i'll write it after this…",{"id":117,"text":118},"38255202","will write an essay about it later, but servers exploding even with experimental alpha users. I'm excited about what H200s or B100s will mean for technology like this one.",{"commenters":108,"max_depth":34,"maker_replies":108,"first_comment_mins":28,"lifespan_hours":34},[121,123,125,126,128,130,132,134,137,139,141,143],[42,28,28,122,43,28],"commenter\u002Fneutral\u002Fperformance",[49,28,34,124,50,28],"project\u002Fneutral\u002Fperformance",[100,28,28,122,43,28],[111,28,28,124,10,127],1,[69,28,28,129,43,28],"commenter\u002Fcritical\u002Fperformance",[53,127,34,131,50,28],"maker\u002Fneutral\u002Fnone",[114,127,28,133,10,127],"maker\u002Fneutral\u002Fexecution",[117,135,28,136,10,127],2,"project\u002Fpositive\u002Fperformance",[92,135,34,138,93,28],"maker\u002Fneutral\u002Fperformance",[140,135,28,138,10,127],"38255701",[80,34,34,142,81,28],"maker\u002Fneutral\u002Fidea",[144,34,28,131,10,127],"38255253",[],[147,152,157,162,166,170],{"id":148,"name":149,"points":150,"category":151},"38243949","dpoint",583,"hardware",{"id":153,"name":154,"points":155,"category":156},"38230325","Spray",393,"dev_tools",{"id":158,"name":159,"points":160,"category":161},"38228788","DataSheetGrid",342,"languages",{"id":163,"name":164,"points":165,"category":161},"38202703","BabaBASIC",234,{"id":167,"name":168,"points":169,"category":156},"38230476","Flyscrape",208,{"id":171,"name":172,"points":173,"category":161},"38276430","Numbat",146,false,"v2",{"id":177,"name":5,"line":5,"ai":174,"cat":18,"year":12,"points":13,"comments":13,"rec":25,"s":26,"top":16,"kw":178,"maker":10,"domain":9,"tier":17,"attempts":28,"strip":179,"url":8,"mix":180,"praise":61,"objection":66,"sub":19},38255201,[19,22,23,24],"030003003030",[28,29,28,28],1794036644644]